Tesis Doctorado Carlos Cobos

MARKOV CHAIN STOCHASTIC SEISMIC INVERSION FOR
ROCK PROPERTIES ESTIMATION
--------------------------------------------------------
A Dissertation Presented to
The Faculty of the Department Earth and Atmospheric Sciences
University of Houston
--------------------------------------------------------
In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy
-------------------------------------------------------
By
Carlos Cobos
December 2014
____________________________________
Carlos Cobos
APPROVED:
____________________________________
Dr. John Castagna, Chairman
____________________________________
Dr. Evgeny M. Chesnokov, Member
____________________________________
Dr. Aibing Li, Member
____________________________________
Dr. Franklin Ruiz, Member
___________________________________
Dean, College of Natural Sciences and

Mathematics
ii
ACKNOWLEDGMENTS
I would like to express my gratitude to John Castagna for being an
outstanding advisor, who has provided assistance, guidance and support. John is
much more than an academic advisor; he is truly a “maestro”. I would also like to
thank Aibing Li, Evgeny Chesnokov and Franklin Ruiz, for their readiness to
serve on my oral exam committee and for their thorough review and constructive
comments on my dissertation and for their tremendous help and support to finish
it. Special thanks go to Aibing Li for her help and guidance as graduate student
advisor in Geophysics.
I want to thank Sylvia Marshal for her patience and help in all these great
years. I am also very grateful to Anne Decour and Hua-wei Zhou; without their
support I would not have been able to complete this dissertation.
I have enjoyed and learned from many discussions with Igor Escobar,
Ezequiel Gonzalez, Andres Mantilla, Per Avseth, Mario Gutierrez, Isao
Takahashi, Manuel Gonzalez and Marcelo Benabentos. I am also very grateful to
Tapan Mukerji and De-Hua Han; I cannot thank them enough for all the support
they have given me all these years.
I have been fortunate to share my personal time with a number of great
friends. Among them the most memorable are Carlos Moreno, Javier Villegas,
Tony De Lilla, Maria Villegas, Alexandra Marquez, Carlos Lopez, Karen Romero,
iii
Laszlo Benkovics, Ana Paredes, Freddy Obregon Roberto Varade and Noemi
Ordonez.
I have infinite gratitude to my parents Gracia and Carlos, my grandfather
Carlos Eduardo, my siblings, Gracia, Carlos Jose, Ligia and Francisco Jesus,
and the rest of my family for their continued trust and support.
Finally, I dedicate this dissertation to my wife Carmen and my children
Alejandra, Carla and Carlos Eduardo. They have been the energy and motivation
that made possible the culmination of my Ph.D.
iv
---------------------------------------------------
An Abstract Dissertation
Presented to
The Faculty of the Department Earth and Atmospheric Sciences
University of Houston
--------------------------------------------------------
In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy
-------------------------------------------------------
By
Carlos Cobos
December 2014
v
ABSTRACT
Our purpose is to introduce a stochastic seismic inversion algorithm based
on the Markov Chain Monte Carlo Simulation. Conventional inversion algorithms
generate elastic properties instead of the key rock properties needed to reliably
characterize the hydrocarbon potential of a given subsurface interval. In contrast,
the inversion scheme presented here generates a set of possible combinations of
rock properties that can explain seismic amplitude responses in terms of
lithology, pore structure and fluid variations. The result of the probabilistic seismic
inversion is a seismic lithofacies catalog that can describe the elastic response of
the studied subsurface interval. The main advantage of this technique is that the
results consist of many equally probable rock property models as an alternative
to multiple elastic property scenarios. Therefore, no post facto elastic-to-rock-
properties conversion is needed. The method might be used either in exploratory
areas or hydrocarbon field development. In exploratory areas, the stochastic rock
physics inversion can support the evaluation for hydrocarbon potential
considering the effects of reservoir properties on seismic signatures for different
geologic scenarios and physical conditions, with the prime goal of minimizing
uncertainties and risk. In field development areas, stochastic seismic inversion
produces multiple equally probable rock property models that can explain the real
3D seismic response; it can be used to constrain possible reservoir models used
for hydrocarbon reserve estimation and reservoir production simulation. The
probabilistic inversion algorithm was successfully tested on a set of real seismic
vi
and well-log data to demonstrate the feasibility of the estimation of critical rock
properties for hydrocarbon exploration, such as total porosity and reservoir
fraction. The real-data test results confirmed the capability of the proposed
inversion technique to accurately predict the rock properties of reservoir seismic
lithofacies, even for seismically thin layers.
vii
CONTENTS
Acknowledgments ...................................................................................... iii

Abstract......................................................................................................vi
List of Tables ............................................................................................. x
List of Figures ............................................................................................xi
1. Introduction ....................................................................................... 1
1.1. Motivation ................................................................................... 1
1.2. Chapter Layout ........................................................................... 4
2. Markov Chain Monte Carlo Simulation.............................................. 7
2.1. Introduction................................................................................. 7
2.2. Random Variables ...................................................................... 7
2.3. Markov Process and Markov Chains ........................................ 12
2.4. Markov Chain Monte Carlo Methods ........................................ 16
2.5. Metropolis–Hastings Algorithm ................................................. 18
3. The Statistical Rock Physics Model ................................................ 22
3.1. Introduction............................................................................... 22
3.2. Rock Physics Models ............................................................... 22
3.3. Deterministic Versus Statistical Models .................................... 24
3.4. Probabilistic Rock Physics Models ........................................... 25
3.5. The multidimensional Non-Parametric Model ........................... 27
3.6. Hybrid Rock Physics Model Example ....................................... 29
3.7. Using MCMCS to Explore Model Space ................................... 36
4. 1D Markov chain stochastic seismic inversion ................................ 38
4.1. Introduction............................................................................... 38
4.2. Proposed Inversion Algorithm .................................................. 39
4.3. Possible Scenario Generation .................................................. 41
4.4. Synthetic Seismic Generation .................................................. 44
4.5. The Misfit Function ................................................................... 48
5. Synthetic Examples ............................................................................. 51
5.1. Introduction ................................................................................... 51
viii
5.2. Synthetic Test 1: Cemented Sand ................................................. 52
5.3. Synthetic Test 2: High Porosity Sands .......................................... 66
5.4. Conclusions................................................................................... 77
6. 1D Real Data Application ................................................................ 79
6.1. Introduction............................................................................... 79
6.2. Seismic Data Conditioning ....................................................... 81
6.3. The Model Space Generation................................................... 83
6.4. The Seismic Stochastic Inversion ............................................. 85
6.5. Analysis .................................................................................... 90
6.6. Conclusions .............................................................................. 95
7. 3D Real Data Application ................................................................ 96
7.1. Introduction............................................................................... 96
7.2. The Kriging Estimator ............................................................... 96
7.3. 3D Real Data Application ......................................................... 99
7.4. Convergence Criteria ............................................................. 109
7.5. Algorithm Efficiency ................................................................ 111
6.7. Conclusions ............................................................................ 112
References ............................................................................................ 113
ix
LIST OF TABLES
Table 2.1: Probability associated to the possible outcomes of random process. .. 9

Table 2.2: Cumulative probability associated to the possible outcomes of the RV
X. ........................................................................................................................ 10
Table 2.3: Markov chain for initial price variation “same”. ................................... 15
Table 2.4: Markov chain for initial price variation “decrease”. ............................. 16
Table 3.1: Vp statistical summary table for seismic lithofacies A and B. ............ 27
Table 3.2: Fluid properties generated using Batzle and Wang (1992) fluid
properties empirical relations. ............................................................................. 32
Table 4.1: Example of the Fisher and Yates shuffle algorithm. The permuted
interval ranges from 1 to 5. The red cells correspond to the number that is
selected in each iteration. ................................................................................... 43
Table 5.1: Comparison between the synthetic data and the depths obtained by
the inversion. ...................................................................................................... 75
Table 6.1: Comparison between the real and the estimated lithology top values
for probability larger than 0.2. ............................................................................. 91
Table 6.2: Comparison between the real and the estimated fluid contact depths.
........................................................................................................................... 93
x
LIST OF FIGURES
Figure 2.1: Probability density function of the discrete random variable described
in Table 2.1. .......................................................................................................... 9
Figure 2.2: Cumulative probability density function of the RV described in Table
2.2. ..................................................................................................................... 11
Figure 2.3: Schematic representation of the transition probability matrix. .......... 15
Figure 2.4: Markov Chain evolution after 500, 1,000, 5,000, 10,000, 25,000 and
50,000 steps. The equilibrium distribution is a Gaussian with zero mean and
variance of one. .................................................................................................. 18
Figure 2.5: Two-dimensional RW with 10,000 points.......................................... 19
Figure 3.1: Deterministic (left) versus stochastic models (right). The stochastic
model generates a set of possible values instead of a simple outcome. ............ 25
Figure 3.2: P-wave velocity histograms for two different lithofacies.................... 26
Figure 3.3: Three-dimensional projections of a ten-dimensional Gaussian PDF
sampled with 1,000 (upper left), 10,000 (upper right), 100,000 (lower left) and
1,000,000 (lower right). ....................................................................................... 29
Figure 3.4: 2D Projections of the multivariate probability distribution function. A)
Bivariate P-wave velocity versus S-wave velocity PDF; B) Bivariate P-wave
velocity versus bulk density PDF; C) Bivariate S-wave velocity versus bulk
density PDF; and D) Bivariate Porosity versus bulk density PDF. ...................... 31
Figure 3.5: 2D Projections of the modeled multivariate probability distribution
function. A) Bivariate P-wave velocity versus Vs PDF; B) Bivariate P-wave
density PDF; and D) Bivariate Porosity versus bulk density PDF. ...................... 33
Figure 3.6: Elastic properties comparison between original saturation (left) and
gas replaced saturation (right). ........................................................................... 34
Figure 3.7: Example of theoretical rock physics models application to measured
P-wave velocity samples. The fluid replacement, cement modeling and porosity
perturbation effects are shown by red, green and black arrows. ........................ 35
P-wave velocity samples. The probability distribution functions for the fluid
replacement, cement modeling and porosity perturbation effects are in red, green
and black, respectively. ...................................................................................... 35
Figure 3.9: Comparison of an original bivariate distribution (left) and the MCMCS
distribution (right). ............................................................................................... 37
xi
Figure 3.10: Mean square error plot for the MCMCS of the example shown in
Figure 3.9. .......................................................................................................... 37
Figure 4.1: The inverse and forward problems are complementary mathematical
problems. ............................................................................................................ 38
Figure 4.2: Simplified 1D stochastic inversion workflow. .................................... 41
Figure 4.3: Schematic representation of the possible scenario generation. ....... 43
Figure 4.4: Comparison among the depth, time and regularized time 1-D P-wave
velocity arrays. The black circles show how the regularization in the time domain
can affect the elastic properties of the modeled subsurface. .............................. 46
Figure 4.5: Schematic representation of the synthetic traces generation
algorithm. ............................................................................................................ 48
Figure 4.6: Misfit calculation example. In this synthetic example the cumulative
L2-norm shows the intervals with larger misfit. .................................................... 49
Figure 5.1: Model 1 synthetic well logs showing (from left to right) facies
indicator, shale volume, P-wave velocity, bulk density, S-wave velocity, porosity,
P-wave impedance and S-wave impedance. ...................................................... 52
Figure 5.2: Synthetic model 1 pore fluids velocity and density profiles. .............. 54
Figure 5.3: Final seismic lithofacies multivariate PDF projections for the Vp-Vs
(Upper), Vp-Rhob (Center), and PhiT-Vp (Lower) spaces. Green, blue, brown
and red contours denote shales, brine sands, oil sands and gas sands,
correspondingly. Some empirical models are plotted for comparison including
Castagna’s mudrock line, Castagna’s 1993, Han’s 1986, Gardner’s sand,
Gardner’s shale, Voigt bound, Reuss bound and Nu’s modified Voigt bound for
quartz mixture. .................................................................................................... 55
Figure 5.4: Model 1 angle dependent reflectivity profiles.................................... 56
Figure 5.5: Ricker wavelet with a central frequency of 25 Hz. ............................ 57
Figure 5.6: Synthetic angle gather generated from the synthetic model 1. ......... 57
Figure 5.7: 2D Projections of the multivariate probability distribution function for
sands (left) and shales (right) calculated from the synthetic model 1. ................ 58
Figure 5.8: Real seismic trace (left) and possible models traces (right)
comparison. The selected traces have L2-norms smaller than 0.9. .................... 59
Figure 5.9: Synthetic versus modeled waveform comparison for four possible
models with α<0.9. ............................................................................................. 60
Figure 5.10: Probable models with α<0.9. The black, red, yellow and green lines
represent the top of upper sand, base of upper sand, top of lower sand and base
of lower sand, respectively. ................................................................................ 61
Figure 5.11: Sandstone fraction for probable models with α< 0.9....................... 62
xii
Figure 5.12: Total porosity of the probable models with α<0.9. The black, red,
yellow and green lines represent the top of upper sand, base of upper sand, top
of lower sand and base of lower sand, respectively. .......................................... 63
Figure 5.13: Volume of shale probable models with α<0.9. The black, red, yellow
and green lines represent the top of upper sand, base of upper sand, top of lower
sand and base of lower sand, respectively. ........................................................ 63
Figure 5.14: Effective porosity probable models with α<0.9. The black, red,
of lower sand and base of lower sand, respectively. .......................................... 64
Figure 5.15: Elastic models with α<0.9. The black, red, yellow and green lines
of lower sand respectively. ................................................................................. 64
Figure 5.16: Normalized histograms comparison between the rock properties of
the synthetic model and the probable solutions, for sand’s shale volume (upper
left), sand’s total porosity (lower left), shale’s shale volume (upper right) and
shale’s total porosity (lower right). Red and blue bars correspond to synthetic
model and solutions, respectively. ...................................................................... 65
Figure 5.17: Probable models with α<0.045. The black, red, yellow and green
lines represent the top of upper sand, base of upper sand, top of lower sand and
base of lower sand, respectively. ........................................................................ 66
Figure 5.18: Model 2 synthetic well logs showing (from left to right) facie
indicator, shale volume, porosity, P-wave velocity, bulk density and S-wave
velocity. Black, blue and red lines correspond to in situ, brine and gas
saturations. ......................................................................................................... 68
Figure 5.19: Synthetic model 2 pore fluids velocity (left) and density profiles
(right). Blue and red lines correspond to brine and gas, respectively. ................ 68
(Upper), Vp-Rhob (Center), and PhiT-Vp (Lower) spaces. Grey, blue, and red
contours denote shales, brine sands and gas sands, correspondingly. Some
empirical models are plotted for comparison, including Castagna’s Mudrock line,
Castagna’s 1993, Han’s 1986, Gardner’s sand, Gardner’s shale, Voigt bound,
Reuss bound and Nu’s modified Voigt bound for quartz mixture. ....................... 69
Figure 5.21: Model 2 angle-dependent reflectivity profiles. ................................ 70
Figure 5.22: Synthetic angle gather generated from synthetic model 2. ............. 71
gas sands (left), shales (center) and brine sands (right), calculated from synthetic
model 2. .............................................................................................................. 71
comparison. The selected traces have L2-norms smaller than 0.88. .................. 72
xiii
models with α<0.88. ........................................................................................... 73
Figure 5.26: Probable models with α<0.88. The green, black and white represent
the shales, gas sands and brine sandstone seismic lithofacies, respectively. .... 74
Figure 5.27: Depth of top of the sand (left), gas-water contact (center) and base
of the sand (right) histograms. ............................................................................ 74
Figure 5.28: Total porosity of the probable models with α<0.88. ........................ 75
Figure 5.29: Volume of shale probable models with α<0.88. .............................. 76
Figure 5.30: Effective porosity probable models with α<0.88 calculated from the
total porosity and shale volume properties. ........................................................ 76
Figure 5.31: Elastic models with α<0.88. ............................................................ 77
Figure 6.1: Well 1 logs for in situ saturation. ....................................................... 80
Figure 6.2: Near angle stack; the top of the upper and lower sands are located at
72 ms and 104 ms, respectively. ........................................................................ 80
Figure 6.3: Angle gather at the Well-1 location (Upper). The amplitude extraction
at the top of the upper sand shows a strong AVO effect (Lower). ...................... 81
Figure 6.4: Well logs converted to time showing the well to seismic tie (right
track); the black line corresponds to the measured seismic trace and blue the
synthetic seismogram. ........................................................................................ 82
Figure 6.5: Angle-dependent extracted wavelets................................................ 83
Figure 6.6: Well 1 logs after fluid substitution. Black, red and blue lines
correspond to in situ, 90% gas and 100% brine-saturated sands. ...................... 84
Figure 6.7: 2D Projections of the multivariate probability distribution function
generated from Well 1 log data and Gassmann fluid substitution. ...................... 84
Figure 6.8: Comparison between real angle traces (upper and lower left) and the
320 solutions for 0° (upper right) and 32° (lower right) incident angles. ............. 86
Figure 6.9: Probable models with α≤0.88. Grey, black and white represent
shales, gas sands and low saturation gas sands, respectively. .......................... 87
Figure 6.10: Total porosity of the 320 probable models. ..................................... 88
Figure 6.11: Volume of shale of the 320 probable models. ................................ 88
Figure 6.12: Effective porosity of the 320 probable models. ............................... 89
Figure 6.13: Possible elastic models with α≤0.88. .............................................. 89
Figure 6.14: Probability of sand vertical plot. ...................................................... 91
Figure 6.15: Total porosity and effective porosity comparison. The black lines
correspond to the well log information and the blue lines to the inverted rock
properties............................................................................................................ 92
xiv
Figure 6.16: Most expected fluid seismic lithofacies (left) and probability of fluid
type (right). ......................................................................................................... 93
Figure 6.17: Comparison between the most expected elastic models (blue line)
against well logs (black lines). ............................................................................ 94
Figure 7.1: Simplified 3D stochastic inversion workflow. .................................. 100
Figure 7.2: Anisotropic semivariogram model with a 0.1 nugget, a sill of 0.9 and
ranges of 500 meters and 50 meters. ............................................................... 100
Figure 7.3: Variance model evolution for one of the stochastic simulation
outcomes. ......................................................................................................... 102
Figure 7.4: Three possible stochastic models. Shales are represented by facies
1, gas sands by facies 2, LSGS by facies 3 and brine sands by facies 4. ........ 102
Figure 7.5: Three possible stochastic total porosity models. ............................ 103
Figure 7.6: Three possible stochastic volume of shale models......................... 103
Figure 7.7: Three possible stochastic P-wave impedance models. .................. 104
Figure 7.8: Three possible stochastic bulk density models. .............................. 104
Figure 7.9: Three possible stochastic S-wave velocity models. ........................ 105
Figure 7.10: Most expected total porosity model generated from 400 stochastic
models. ............................................................................................................. 106
Figure 7.11: Most expected volume of shale model generated from 400
stochastic models. ............................................................................................ 107
Figure 7.12: Most expected volume of P-wave velocity generated from 400
Figure 7.13: Most expected volume of density generated from 400 stochastic
models. ............................................................................................................. 108
Figure 7.14: Most expected volume of S-wave velocity generated from 400
Figure 7.15: Models randomly selected from the real dataset inversion. .......... 109
Figure 7.16: Comparison between real angle trace (right) and 150,000 solutions
(left) randomly selected from the five million realizations performed for the real
example inversion............................................................................................. 110
Figure 7.17: Algorithm efficiency, the proposed algorithm increase linearly with
the number of tested models. ........................................................................... 111
xv
1. INTRODUCTION
1.1. Motivation
Conventional seismic reservoir characterization (SRC) techniques were
developed more than forty years ago for exploration plays where hydrocarbon
seismic responses were relatively easy to identify. Early reservoir
characterization workflows were usually based on direct hydrocarbon indicator
(DHI) identification techniques centered on poststack seismic amplitude analysis
and amplitude versus offset (AVO) inversion. DHIs plays are usually related to
shallow high porosity reservoirs with significantly lower acoustic impedance
compared to the surrounding rocks. The associated seismic signatures of these
hydrocarbon-filled high porosity reservoirs can be anomalous high amplitude
reflections called “bright spots”.
Nowadays, conventional seismic-reservoir-characterization techniques are
becoming obsolete, since the oil industry is moving to explore areas in which the
hydrocarbons are located in deeper and more-complex reservoirs. These new
hydrocarbon plays are characterized by low-porosity and low-permeability
reservoirs with nearly undetectable pore fluid response. This means that the
future seismic-reservoir characterization goal is to predict rock properties, such
as the porosity, lithology and rock fabric of compacted and cemented porous
rock.
1
The second-most important SRC challenge is to improve seismic vertical
resolution. Currently, seismic inversion resolution is still low in new
exploration/development challenges and the improvement of seismic derived
elastic parameters is paramount for the application of reservoir properties
prediction. Techniques such as stochastic inversion have been initially developed
in an attempt to obtain from seismic data quantitative information about
subsurface rock properties on a very detailed scale.
The main limitation of current seismic-inversion algorithms is that the
solutions are usually limited to the acoustic (P-wave impedance) or elastic (P-
wave impedance, S-wave impedance, and bulk density) properties of the studied
subsurface intervals. However, these properties are not directly useful for
characterizing hydrocarbon potential. To accurately characterize the oil potential
of a possible of subsurface reservoir, we need to know key rock properties, such
as lithology, porosity and fluid type.
The goal of this dissertation is to introduce an inversion technique based
on the Markov Chain Monte Carlo Simulation (MCMCS) that can be implemented
in a stochastic petrophysical inversion scheme. Our proposed stochastic seismic
inversion’s objective is to produce a set of probable rock property volumes that
can describe the elastic seismic response of the studied interval and its
associated uncertainties.
2
The main advantage of this petrophysical inversion technique is that the
results are multiple-probable-rock-property models instead of multiple-elastic-
property scenarios. Therefore, no post facto elastic-to-rock properties conversion
is needed. The method might be used either in exploratory areas or hydrocarbon-
field development. In exploratory areas, the proposed stochastic inversion
algorithm can support the evaluation for hydrocarbon potential considering the
effects of reservoir properties on seismic signatures for different geologic
scenarios and physical conditions, with the prime goal of minimizing uncertainties
and risk. In field development areas, the stochastic inversion produces many
equally probable rock property models that can explain the real 3D seismic
response; it can be used to constrain possible reservoir models used for
hydrocarbon reserve estimation and reservoir production simulation.
In this dissertation we selected a probabilistic solution to tackle the
problem of estimating the physical properties of subsurface rocks based on
surface seismic data. According to Tarantola (2005), most probabilistic inverse
problems can be resolved using the following equation:
( ) ( ) ( ( )) (1.1)
where ( ) is the a posteriori probability density function in the model
space, K is a normalizing constant, ( ) is the a priori probability density
function in the model space, ( ) is the data probability density function and
( ) is the forward modeling operator that relates the model space to the data
3
space. The operator ( ) is usually non-linear and its role is to predict the
values of the observable parameters that would correspond to a given
model . This theoretical prediction can be described as follows:
( ) (1.2)
The main contributions of this dissertation are: 1) the introduction of the
Metropolis–Hastings MCMCS algorithm to explore multivariate non-parametric
statistical rock physics models; and, 2) the introduction of an inversion technique
that directly produces rock property volumes instead of elastic property volumes;
hence, the inversion results can be directly introduced into the geological model
of the studied area.
1.2. Chapter Layout
Chapter 2 introduces the basis of the Markov Chain Monte Carlo
Simulation (MCMCS) and its application to the proposed stochastic inversion.
First, we introduce some basic concepts of descriptive statistics. Then we define
Markov Chains and their application to Monte Carlo simulations. Finally, it is
shown how the Metropolis–Hastings MCMCS algorithm is used to explore the
multidimensional solution space in a much more effective way than systematic
sampling algorithms.
Chapter 3 focuses on the probabilistic rock physics model used as model
space for the proposed inversion scheme. We show how probability density
4
functions are generated from the integration of the available well and geological
data, and are supported by theoretical rock physics models. Second we describe
how the MCMCS is used to populate the model space for the proposed
stochastic seismic inversion.
Chapter 4 presents our novel 1D stochastic seismic inversion algorithm
based on the Metropolis–Hastings MCMCS algorithm. The inversion workflow
and its three main steps are explained: the rock properties scenario generation,
the synthetic data generation and the comparison between the real and the
synthetic traces. We illustrate how the Metropolis–Hastings MCMCS algorithm is
used to populate the rock properties of the tested models.
Chapter 5 presents two synthetic examples to illustrate the proposed
inversion technique. We show the results of the application of the proposed
inversion scheme on two different sets of synthetic data. The synthetic data sets
were generated based on real well data, with the goal of testing the inversion in
two different geological environments. The first synthetic example represents a
geological setting where the sands have larger impedance than that of the
surrounded shales. In the second example, we tested the algorithm in a high
porosity environment where the fluid property variations have a very strong effect
on the elastic properties of the studied reservoirs.
In Chapter 6, we present the first 1D application of the proposed inversion
scheme on a real set of data that include seismic and well log information. The
5
target interval is made up of four seismic lithofacies: shales, gas-filled sands, low
saturation gas sands (LSGS) and brine sands. The objective of this inversion is
to predict total porosity, volume of shale, effective porosity and fluid type
distribution in the studied interval.
Chapter 7 presents the 3D inversion of the real data set introduced in
Chapter 6. The 3D algorithm integrates the 1D stochastic inversion procedure
with a Kriging engine to produce a lateral constraint rock property model. We
analyzed the same data set used in Chapter 6 that corresponds to tertiary marine
sediments deposited on a platform margin.
6
2. MARKOV CHAIN MONTE CARLO SIMULATION
2.1.Introduction
This chapter introduces basic concepts of statistics and the Metropolis–
Hastings Monte Carlo Markov Chain Simulation (MCMCS) algorithm. This
theoretical introduction is the foundation of the seismic inversion technique based
on stochastic simulation proposed in this dissertation. The MCMCS algorithm is
used to explore the multidimensional solution space in a much more effective
way than systematic sampling algorithms. Its application is especially appropriate
in inverse Earth science problems, in which the number of variables tends to be
very large. Before introducing the MCMCS algorithm we define some basic
statistical concepts in Section 2.2. These basic concepts are random variables,
probability density functions, cumulative density functions and the most important
descriptive statistical parameters.
2.2. Random Variables
A random variable (RV) is a mathematical variable whose possible values
are the numerical outcomes of a random process. According to Papoulis
(Papoulis 1984), any function g(x) of real variables can be a RV if g(x) has the
following properties:
a) Its domain must include all the range of RV x,
b) The set {X<=x} is an event for any real number x, and
7
c) The probability of the events {X=±∞} must be zero.
Random variables can be discrete or continuous. RVs that are continuous
can assume any value within a given range. For instance, if we randomly select
several individuals from a population and measure their height, the results can
take on any value within a given range, maybe from 2 to 7 feet. In contrast,
discrete variables can only assume certain values within a defined range. For
example, when we flip a coin and count the possible outcomes, no matter the
number of flips, the possible outcomes can only be heads or tails. Because the
number of heads or tails results from a random process flipping a coin is a
discrete RV.
Random variables are defined by their probability density function (PDF)
that describes the likelihood for a given RV to take on a given value. The
probability of an RV taking a particular range of values is defined by the integral
its PDF taking over the analyzed range.
The following example will illustrate some concepts related to the random
variables. Assume an RV X that can only take the values 1, 2, 3, 4 or 5. The
probabilities associated with the possible outcomes are shown in Table 2.1.
8
Table 2.1: Probability associated to the possible outcomes of random process.
Outcome Probability
1 0.05
2 0.25
3 0.4
4 0.25
5 0.05
The probability that X is equal to 1 or 2 is the sum of two probabilities: P(X
= 1 or X = 2) = P(X = 1) + P(X = 2) = 0.05 + 0.25 = 0.3. Correspondingly, the
probability that X is greater than 1 is equal to 1 - P(X = 1) = 1 - 0.05 = 0.95. The
set of possible outcomes and their associated probabilities shown in Table 2.1
are the PDF of the RV X (Isaaks and Srivastava, 1989); Figure 2.1 shows the
PDF plot of the RV shown in Table 2.1.
0.4
0.35
0.3
0.25
Probability
0.2
0.15
0.1
0.05
0
1 2 3 4 5
Outcome
Figure 2.1: Probability density function of the discrete random variable described
in Table 2.1.
9
Suppose the random variable X may take N different values; the PDF of
the discrete RV X is then defined as:
( ) . (2.1)
The PDF of a discrete RV must satisfy the following:
a) 0 < Pi< 1 for each i

b) ∑ .
By definition, the integral of the PDF (the sum of discrete RVs) is called
the cumulative distribution function (CDF). As mentioned, the CDF is defined as
the probability that a RV X has a value less than or equal to x. The CDF of a
discrete random variable X is related to the PDF by the following expression:
{ } ∑ (2.2)
Table 2.2 shows the CDF of the RV X shown in Table 2.1. The CDF of the
RV X is displayed in Figure 2.2.
Table 2.2: Cumulative probability associated to the possible outcomes of the RV

X.
Outcome CDF
1 0.05
2 0.3
3 0.7
4 0.95
5 1
10
1
0.9
0.8
Cummulative Probability
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1 2 3 4 5
Outcome
Figure 2.2: Cumulative probability density function of the RV described in Table

2.2.
When the PDF of an RV is known, we can calculate many parameters that
are useful to describe the statistics of the analyzed random variable. These
parameters include expected value and standard deviation of the RV. The
expected value of an RV X is defined as the weighted average of all the
outcomes; each one is weighted by the probability of occurrence of that outcome:
{ } ∑ (2.3)
Based on the information shown in Table 2.1 we can calculate the
expected value of the RV X as:
{ }
Another useful parameter to describe the statistics of a RV X is variance.
The variance of a RV is a measure of the distance from the outcomes to the
11
expected value, more commonly known as dispersion. A small variance indicates
that the outcomes are very close to the expected value. In contrast, a large
variance indicates that the data is very spread out around the mean. Variance
can be estimated as:
{ } ∑ (∑ ) . (2.4)
2.3. Markov Process and Markov Chains
As we introduced in the previous section, a random or stochastic process
is a sequence of outcomes that depend on some probability. A Markov process is
a stochastic “memory-less” process in which predictions of future outcomes of
the stochastic process are based only on its present state. A random variable is
defined as a Markov process if the RV has the following properties:
a) The number of possible outcomes or states is finite;
b) The outcome at any stage depends only on the outcome of the previous
stage; and
c) The probabilities are constant over time.
Consequently a Markov process can be defined as follows:
( | ) ( | ) (2.5)
A sequence of random variables generated by a Markov process is called
a Markov Chain. Any Markov Chain is defined by the transition probability to
12
move from one stage to another. The probability that an RV in state xi will change
to a state xj in a single step is defined as:
( ) ( | ). (2.6)
The probability that the chain is in state xi at step n can be calculated
using the Chapman-Kolmogorov equation that is the sum of the probability of
being in a particular state at the current step and the transition probability from
the current state to xj:
( ) ∑ ( | ) ( ). (2.7)
A more convenient way to express the Chapman-Kolmogorov equation is
the matrix form:
( ) ( ) , (2.8)
where S(n) is the probability vector at the nth step, S(0) is the probability vector at
the initial step, and P is the probability transition matrix.
A Markov Chain can be illustrated using a simple example. Suppose that the
price variation of a particular share in the stock market can be predicted using a
Markov process and the possible price states are: “decrease”, “same” or
“increase”. Therefore, the price of the share tomorrow only depends on today’s
13
share price. Assuming that today the price variation is “same”, the transition
probability matrix (Figure 2.3) can be written as follows:
( ).
The initial state vector is:
( ) ( ).
Tomorrow’s price variation will be:
( ) ( )( ) ( ).
The sixth day’s price variation will be:
( ) ( )( ) ( )
14
Figure 2.3: Schematic representation of the transition probability matrix.
The summary of the Markov Chain for the first six days is shown in Table
2.3. However, if the initial price variation is “decrease” instead of “same” (see
Table 2.4), the Markov Chain reaches the same probability values at the sixth
day. Hence, on the sixth day, the Markov Chain reaches its stationary
distribution, where the probability values are independent of the initial state.
Table 2.3: Markov chain for initial price variation “same”.
Day Decrease Same Increase
1 0.2500 0.5000 0.2500
2 0.3000 0.4000 0.3000
3 0.3100 0.3800 0.3100
4 0.3120 0.3760 0.3120
5 0.3124 0.3752 0.3124
6 0.3125 0.3750 0.3125
15
Table 2.4: Markov chain for initial price variation “decrease”.
Day Decrease Same Increase
1 0.4000 0.3000 0.3000
2 0.3250 0.3600 0.3150
3 0.3145 0.3720 0.3135
4 0.3129 0.3744 0.3128
5 0.3126 0.3749 0.3126
6 0.3125 0.3750 0.3125
2.4. Markov Chain Monte Carlo Methods
Monte Carlo Simulation (MCS) methods are a set of computational
algorithms used to solve numerical problems based on random numbers. One of
the most common applications of MCS is to calculate the probability of a specific
outcome by randomly sampling a large number of scenarios. MCS methods
have many applications in physics, finance, engineering, and mathematics. In
this dissertation we use the MCS to explore the space of possible solutions used
in the proposed seismic inversion algorithm.
Markov Chain Monte Carlo (MCMC) is a family of algorithms for sampling
desired probability distributions based on Markov processes. The shape of the
desired distribution has to be similar to the MCMC equilibrium distribution; this
means that the quality of the MCMC distribution improves with the number of
steps. A Markov Chain is an ideal method for Monte Carlo simulations due to the
16
memoreless nature of Markov stochastic processes. This Monte Carlo method
combines all the desired features for stochastic simulation:
a) It is simple and easy to construct a Markov Chain with the desired
properties;
b) It generates statistically consistent samples from the target distribution;
and
c) The samples drawn are dependent and follow a Markovian framework
which allows more robust convergence.
Markov Chain mixing time is by definition the number of steps needed to
reach the equilibrium with tolerable error. The better the Morkov Chain, the faster
the mixing time. Figure 2.4 shows an example of a Markov Chain with a
Gaussian as steady distribution; in this example, the desired distribution is
reached after 50,000 steps.
17
500 Samples 1000 Samples
40 60
50
30
40
Counts
20 30
20
10
10
0 0
-4 -2 0 2 4 -4 -2 0 2 4

300 600
250 500
200 400
Counts
150 300
100 200
50 100
0 0
-4 -2 0 2 4 -4 -2 0 2 4

1500 4000
3000
1000
Counts
2000
500
1000
0 0
-4 -2 0 2 4 -4 -2 0 2 4
Property Property
Figure 2.4: Markov Chain evolution after 500, 1,000, 5,000, 10,000, 25,000 and
50,000 steps. The equilibrium distribution is a Gaussian with zero mean and
variance of one.
2.5. Metropolis–Hastings Algorithm
A random walk (RW) is a mathematical stochastic path that is generated
from a series of random steps. RW theory plays an important role in physics,
statistics and engineering, since many physical phenomena can be modeled
using a stochastic method; coin tosses, Brownian motion and stock prices are
examples of processes that can be modeled as RWs. These methods are
particularly useful as sampling technique since RW sampling methods are least
18
biased compared to systematic sampling algorithms. However, RWs can lead to
poor representation of the real distribution in very-large-dimension problems.
Figure 2.5 shows a two-dimensional RW with 5,000 steps. In a simple
RW, the location can only jump to neighboring sites of the lattice. In a simple
symmetric RW on a locally finite lattice (Figure 2.5), the probabilities of the
location jumping to each one of its immediate neighbors are the same.
Figure 2.5: Two-dimensional RW with 10,000 points.
RW algorithms are generally based on Markov Chain processes. In this
dissertation, we implemented a Markov Chain Monte Carlo method called the
Metropolis–Hastings algorithm. The Metropolis–Hastings algorithm was originally
introduced in 1953 by a group of scientists (Metropolis et. al., 1953) from Los
19
Alamos National Laboratory, and extended to its general form by W. K Hasting in
1970 (Hastings, 1970). The idea of the Metropolis–Hastings Monte Carlo
algorithm is to compute values of the function P(x) that are proportional to a
known probability distribution F(x) P(x)=K*F(x), where the proportional constant K
is unknown and very difficult to estimate. The Metropolis algorithm generates a
sequence of draws as follows:
1. Randomly select an initial value x0 that satisfied the relation F(X0)>0.
2. Using the current value (xt) draw, generate the candidate value (xt+1) from
a proposed transition probabilistic function q(xt,xt+1).
3. From the candidate (xt+1) and the current (xt) points, calculate the ratio of
the density at the candidate and current points:
( ) ( )
; (2.9)
( ) ( )
If α≥1 accept the candidate point (xt+1) as current value and return to step
2; otherwise, accept the candidate with probability α, else reject it and return to
step 2.
Finally, the Metropolis–Hastings algorithm can be written as follows:
( ) ( )
( ). (2.10)
( ) ( )
20
As with any MCMC methods, the Metropolis–Hastings algorithm
approaches the desired distribution as the number of steps increases. A key
issue in the successful implementation of the Metropolis-Hastings or any other
MCMC sampler is the number of runs (steps) until the chain approaches
stationarity (the length of the burn-in period). Typically, the first 1,000 to 5,000
elements are thrown out, and then one of the various convergence tests is used
to assess whether stationarity has indeed been reached.
A poor choice of starting values and/or proposal distribution can greatly
increase the required burn-in time, and an area of much current research is
whether an optimal starting point and proposal distribution can be found. One
suggestion for a starting value is to start the chain as close to the center of the
distribution as possible, for example by taking a value close to the distribution’s
mode. A chain is said to be poorly mixing if it stays in small regions of the
parameter space for long periods of time, as opposed to a well-mixed chain that
seems to effortlessly explore the space.
21
3. THE STATISTICAL ROCK PHYSICS MODEL
3.1.Introduction
This chapter introduces the concept of the statistical rock-physics model
and illustrates how well-log data and theoretical rock-physics models are used to
populate the model space of the proposed seismic-inversion algorithm.
Sedimentary rocks are aggregates of compacted crystalline grains and
organic matter that are often amalgamated by cement minerals. The free space
between the mineral grains is often called intergranular space or porosity. In in
situ conditions, the intergranular space is filled by liquid or gaseous fluids. The
solid portion of the rock is known as the rock’s matrix, or more commonly, “the
matrix”. The rock’s matrix geometry or how the crystalline grains are arranged in
a sedimentary rock is usually extremely complex and depends on many
variables, such as provenance, sorting, burial history, diagenesis, etc. We need
to define an enormous number of parameters to accurately relate the geometry
of most sedimentary rocks and their elastic moduli; among these parameters, we
have volume fractions of minerals, nature of boding on the grain interface, type
and volume fraction of cements, etc. (Ruiz, 2009).
3.2. Rock Physics Models
Rock-physics models are mathematical models that are related to rock
properties (e.g., porosity, mineral and cement composition pore geometry, elastic
22
moduli, etc.). Rock-physics models are used to estimate rock properties that
have not been sampled. Most rock-physics models can be classified into three
groups: theoretical, empirical, and heuristic (Avseth et al., 2005).
According to Avseth (2005), theoretical models are continuum mechanics
approximations of the elastic, viscoelastic, or poroelastic properties of rocks. Biot
theory (Biot, 1956) is one of the first to model the coupled mechanical behavior of
a porous rock filled with a linearly viscous fluid. At zero frequency, the Biot
relations turn into the Gassmann equation (Gassmann, 1951).
Empirical models are based on the statistical regression analysis of
experimental observations. Most empirical models are only valid for the specific
data sets and physical conditions where the data was measured. Empirical rock
physics models are widely used to predict the properties of rock for which
measured data is not available. Examples of empirical models are Castagna’s
mudrock line (Castagna et. al., 1985), Gardner’s density relation (Gardner et. al,
1974) and Han’s Gulf of Mexico relations (Han, 1986).
Heuristic rock-physics models are mathematical relations based on the
experience-based techniques for problem solving. Since these models are non-
mathematically rigorous, they are not optimal. Popular examples of heuristic
models are the Wyllie time-average equation (Wyllie et. al,. 1956) and the
Greenberg and Castagna relations (Greenberg and Castagna, 1992).
23
3.3. Deterministic Versus Statistical Models
A deterministic model is a mathematical relation in which outcomes are
precisely determined through known relationships among variables. In such
models, a given input will always produce the same output. In deterministic
models the estimation error is assumed to be zero or negligible. Unfortunately, in
Earth science it is impossible to gather the required knowledge to create a
deterministic model that accurately relates the physical properties of most
sedimentary rocks. In contrast, statistical models generate a set of possible
values with their associated probability of occurrence (Figure 3.1). Therefore,
theoretical and empirical rock physics models can be considered deterministic
and statistical models respectively. While deterministic models rely on the
physics of the phenomenon, probabilistic models trust in an adequate statistical
sampling of the analyzed set of data. It is important to emphasize that a poor
sampling of the data space will lead to the formulation of an inaccurate or
incorrect probabilistic model.
24
Figure 3.1: Deterministic (left) versus stochastic models (right). The stochastic
model generates a set of possible values instead of a simple outcome.
3.4. Probabilistic Rock-Physics Models
Probabilistic rock physics models are physical models that relate the rock
properties of a given seismic lithofacies using probabilistic techniques. Avseth
(2000) defined a seismic lithofacies as a seismic-scale sedimentary unit that is
characterized by its lithology, bedding configuration, petrography and elastic
properties. Seismic lithofacies can be used to estimate the statistical parameters
of the rocks that constitute the sedimentary unit. Since these rocks have similar
depositional and diagenesis history, we can assume the stationarity of the
statistics of the rock’s properties for any particular seismic lithofacies (Avseth,
2000).
25
We usually observe data scatter when we plot elastic properties against
rock properties for a given seismic lithofacies. Due to data scatter, descriptive
statistics are often used to summarize the elastic properties of a given seismic
lithofacies. The descriptive statistics is useful to define the most important
properties of the statistical populations. Figure 3.2 shows the P-wave velocity
histograms of two different lithofacies; the histograms clearly show that
lithofacies A has a lower most expecting value and larger velocity dispersion than
lithofacies B. Descriptive statistical measurements are normally used to define
the tendency of the data (Table 3.1), for example, when most of the data is
gathered or gives an idea of the data dispersion.
Figure 3.2: P-wave velocity histograms for two different lithofacies.
26
Table 3.1: Vp statistical summary table for seismic lithofacies A and B.
Seismic Lithofacies A Seismic Lithofacies B

Mean 3849.9 4553.2
Median 3831.4 4568.4
Standard Deviation 187.85 100.31
Variance 35288 10063
For most practical applications, descriptive statistics is not enough to
represent the complexity of the relations among rock properties, since the
definition of central estimators (like the mean or the median) and the dispersion
estimators (like the variance and standard deviation) generally do not have clear
meaning in multidimensional non-parametric spaces. Hence, to avoid the
limitations of descriptive statistics, non-parametric probability density functions
(PDFs) are introduced to better describe the dependence among rock properties,
including its non-uniqueness and non-linearity (Takahashi, 2000). Usually, PDFs
are generated from the integration of the available well and geological data and
are supported by theoretical rock physics models. In fact, multivariate PDFs are
usually hybrid rock physics models that can describe measured rock samples
and scenarios not sampled by measured data.
3.5. The multidimensional Non-Parametric Model
A poorly mixed chain can arise because the target distribution is complex
(multi-modal) and our choice of starting values traps the Markov Chain to one of
27
the modes. An approach to solve this problem is to transform the complex target
distribution to a simpler parametric distribution like a Gaussian distribution, and
then proceed with the MCMCS. This process is called normal scoring and is
commonly used in many statistical applications. Normal scoring processes can
be applied to any number of dimensions. However, it is easy to underestimate
the sampling (leaving many empty zones) of large-dimensional inverse problems.
To solve this issue, the number of samples has to be largely increased with the
number of dimensions. Figure 3.3 shows four three-dimensional projections of a
ten-dimensional Gaussian distribution sampled with 1,000, 10,000, 100,000 and
1,000,000 draws. We can see that the distribution sampled with 1,000 draws
does not represent the shape of the target Gaussian distribution. The modeled
distribution approaches the target distribution when the number of samples
exceeds 100,000 draws. After the MCMCS is performed, the resulting normal
Gaussian distribution is converted back to its original shape, applying reverse
normal scoring.
28
Figure 3.3: Three-dimensional projections of a ten-dimensional Gaussian PDF
sampled with 1,000 (upper left), 10,000 (upper right), 100,000 (lower left) and
1,000,000 (lower right).
3.6. Hybrid Rock-Physics Model Example
In hydrocarbon exploratory areas where the availability of well log data is
usually limited, rock physics modeling is used to populate the multivariate PDFs
with data or seismic lithofacies that are not sampled by the existing wells, but are
probable in the studied area (Gonzalez, 2006). An example of a hybrid rock
physics model can be made by the application of Gassmann’s equation
(Gassmann, 1951) to a set of well log data. Gassmann’s equation allows us to
29
estimate the effective bulk modulus of a fluid filled porous rock when the original
fluid is replaced by a second fluid. This is a very important problem in
hydrocarbon exploration, where fluid replacement is commonly used to sample
different fluids and saturations scenarios in potential reservoir rocks. Gassmann’s
equation is written as follows:
( )
(3.1)
, (3.2)
where:
= Effective bulk modulus of the rock with pore fluid,
= Effective bulk modulus of the dry rock,
= Bulk modulus of the pore grain material,
= Effective bulk modulus of the fluid,
= Porosity,
= Effective shear modulus of the rock with pore fluid,
= Effective shear modulus of the dry rock.
Some of the projections of a multivariable probability density function are
shown in Figure 3.4. The shown PDF was generated from well log data and it
defines the relations among the rock properties for clean sandstone. This
sandstone represents a lithofacies whose intergranular space is filled by brine.
30
Figure 3.4: 2D Projections of the multivariate probability distribution function. A)
Bivariate P-wave velocity versus S-wave velocity PDF; B) Bivariate P-wave
density PDF; and D) Bivariate Porosity versus bulk density PDF.
We applied the Gassmann relations to predict the sandstone elastic
properties when the pores are filled with gas. We estimated the acoustic
properties of the brine and gas using the Batzle and Wang empirical model
(Batzle and Wang, 1992) and assuming salinity of 15,000 ppm, temperature of
84°C, pore pressure of 25 MPa and gas specific gravity of 0.6. The irreducible
water saturation of the modeled sandstone was assumed 8%; hence the fluid
properties were calculated for a saturation of 92% gas and 8% brine using the
31
homogeneous saturation model (Mavko et al., 1998). The fluid properties at the
specified temperature and pressure are displayed in Table 3.2.
Table 3.2: Fluid properties generated using Batzle and Wang (1992) fluid
properties empirical relations.
Brine Gas 90% Gas and 10% Brine
Bulk Modulus [GPa] 2.59 0.05 0.06
Density [g/cc] 0.99 0.126 0.21
Vp [km/sec] 1.61 0.63 0.53
The three-dimensional projection of the multivariate probability distribution
function of the modeled gas saturated sandstone is shown in Figure 3.5. As is
expected, the replacement of the original fluid with less dense and more
compressible fluid will decrease the density and bulk modulus. In contrast, the S-
wave velocity slightly increases due to the density effect on the S-wave velocity
expression. The comparison of the elastic properties histograms for the original
and fluid replaced sands shows that the fluid substitution introduces more
uncertainty in the gas filled rock in comparison with the brine-filled rock (see
Figure 3.6). A similar observation was made done by Sengupta (2000) in her
sensitivity analysis of fluid replacement.
32
Figure 3.5: 2D Projections of the modeled multivariate probability distribution
function. A) Bivariate P-wave velocity versus Vs PDF; B) Bivariate P-wave
density PDF; and D) Bivariate Porosity versus bulk density PDF.
33
Figure 3.6: Elastic properties comparison between original saturation (left) and
gas replaced saturation (right).
This probabilistic rock physics modeling process can include any simple or
complex rock physics model as long as the rock property can be defined in the
multivariate PDF. For instance, the Dvorkin cemented sand model (Dvorkin and
Nur, 1996) can be used to theoretically add cement to sand measurements.
Other applications of theoretical rock physics models are shown in Figure 3.7.
The effect of fluid replacement, adding cement or reducing porosity over the P-
wave velocity (Vp) are displayed in the porosity-Vp space. Figure 3.8 shows the
probability functions for the modeled seismic lithofacies.
34
P-wave velocity samples. The fluid replacement, cement modeling and porosity
perturbation effects are shown by red, green and black arrows.

P-wave velocity samples. The probability distribution functions for the fluid
replacement, cement modeling and porosity perturbation effects are in red, green
and black, respectively.
35
3.7. Using MCMCS to Explore Model Space
As we have mentioned, our proposed stochastic inversion scheme relies
on a Markov Chain Monte Carlo Simulation (MCMCS) engine to explore the
model space represented by a non-parametric multivariate probability density
function. The use of Monte Carlo simulation is central to the resolution of inverse
problems with large-dimensional model spaces for which systematic evaluation
of model spaces on a regular grid is not feasible (since the required number of
points to evaluate is too large).
When the MCMCS algorithm is well implemented, the shape of the target
PDF does not change with the number of simulated samples; this means that the
MCMCS is really random and unbiased. An example of the MMCMCS
implementation is shown in Figure 3.9. The shape of the simulated distribution is
very similar to the original PDF, even though the number of simulated points is
more than one hundred times the number of original data points. We use the
Mean Squared Error (MSE) between the simulated and target distributions to
estimate the optimum number of steps to reach MCMCS equilibrium.
Figure 3.10 shows the MSE plot for the MCMCS example shown in Figure
3.9. For this particular MCMCS implementation the equilibrium is reached after
8,000 steps. The MSE is estimated as follows:
( ) ],
Where = Target Distribution and =Simulated Distribution.
36
Figure 3.9: Comparison of an original bivariate distribution (left) and the MCMCS
distribution (right).
Figure 3.10: Mean-square-error plot for the MCMCS of the example shown in
Figure 3.9.
37
4. 1D MARKOV CHAIN STOCHASTIC SEISMIC INVERSION
4.1. Introduction
This chapter presents the 1D Markov Chain stochastic inversion algorithm. First,
it shows the general workflow for the 1D inversion scheme. Second, it describes
in detail all the steps of the general 1D workflow.
In physics the inverse problem consists of using measurements to infer
the parameters that describe a system (Tarantola, 2005). In contrast, the forward
problem involves the prediction of the measurements given the parameters that
describe the studied system (Figure 4.1). In Earth science the goal of the inverse
problem is to estimate the earth’s properties from a series of geophysical
observations.
Figure 4.1: The inverse and forward problems are complementary mathematical
problems.
38
The content of this chapter was partially presented in the paper IPTC-18040 of the 2014 International Petroleum
Technology Conference in Kuala Lumpur.
This dissertation is focused on the specific problem of estimating the
earth’s parameters from surface seismic data. As with any inverse problem in
earth science, seismic inversion does not have a unique solution. Hence, we
proposed a novel inversion scheme that is based on a probabilistic method
instead of a deterministic approach. When the inverse problem is solved using a
probabilistic approach the system parameters are represented by a PDF
generated from the available data and model. The probabilistic theory applied to
this dissertation can be applied in any kind of inverse problem, including
nonlinear problems. To solve most probabilistic inverse problems, we need to
implement an extensive exploration of the system space. Due to the
multidimensional nature of the Earth science problems, we cannot perform a
systematic system space exploration. Instead, we perform possible solution
space exploration using Markov Chain Monte Carlo methods.
4.2. Proposed Inversion Algorithm
The proposed 1D stochastic inversion algorithm is based on a Markov
Chain stochastic simulation that uses statistical rock physics to constrain the
space of possible solutions. Mukerji et al. (2001) introduced the statistical rock-
physics methods as a way to combine rock physics, information theory, and
geostatistics to reduce uncertainty in reservoir property estimation. Our
fundamental contribution is using the Metropolis–Hastings MCMCS algorithm to
generate multivariate non-parametric statistical rock physics models and
introducing the inversion technique that directly produces rock property volumes
39
instead of elastic property volumes; using our methods, inversion results can be
directly introduced into the geological model of the studied area.
Figure 4.2 shows an idealized scheme for a 1D inversion algorithm. The
inversion technique consists of four main steps: the rock properties scenario
generation, the synthetic data calculation and the comparison between the real
and synthetic traces. In each realization, a model of the subsurface properties is
generated, and the model properties are calculated from the previously defined
multivariate PDF. After the subsurface model is generated, a synthetic trace is
produced using the elastic properties from the subsurface model. Finally, the
synthetic trace is compared to the real seismic trace; if the similarity between
both traces is larger than a given threshold, the model is accepted as a possible
solution and stored. In contrast, if the likeness between the real and the synthetic
traces is smaller than the mentioned threshold, the model is erased and a new
model is generated.
An important characteristic of this proposed inversion scheme is that the
model is generated in a depth domain, so there is no need to convert the model
to depth after a probable subsurface model has been accepted. A detailed
description of each inversion step is shown in the Chapter 5. The proposed
algorithm produces rock property traces instead of elastic property traces.
Therefore, the inversion outcome can be directly introduced into the geological
model of the studied area.
40
Figure 4.2: Simplified 1D stochastic inversion workflow.
4.3. Possible Scenario Generation
Every 1D model is represented by two main attributes. The first is the
layers geometry or how the seismic lithofacies are arranged within the studied
interval. Each seismic lithofacies is defined as a discrete variable with a related
rock property distribution. The second is the rock properties of the analyzed
lithofacies. The scenario generator algorithm works on two levels. First the
routine generates the geometry of the probable earth models. Once the geometry
41
is built, we assign rock properties to each layer. The rock properties are assigned
by using a Markov Chain-based stochastic simulation algorithm. This process is
illustrated in Figure 4.3. First, the fraction of the seismic lithofacies that constitute
the analyzed interval is generated. Once the 1D array with the fractions of the
seismic lithofacies that constitute the study interval are generated, the 1D array
with the distribution of these lithofacies is produced using the Fisher-Yates
shuffle algorithm. This algorithm was originally described by Ronald A. Fisher
and Frank Yates in 1938 (Fisher et al., 1938). The Fisher-Yates shuffle algorithm
is a numerical method for generating random permutations of a finite number of
groups. The most important property of the Fisher–Yates shuffle is that it is
unbiased; consequently, every permutation is equally probable. The algorithm to
generate a random permutation of the numbers one through N goes as follows:
a) Write down the numbers from one through N;
b) Pick a random number X between 1 and the number of remaining
unchosen numbers;
c) Counting from the low end, strike out the Xth number not yet selected and
store it;
d) Repeat from step 2 until all the numbers have been selected.
The 1D array of numbers stored after all the numbers are selected is now
a random permutation of the original numbers. If the random number generator is
truly random and unbiased, the resulting permutation will be also.
42
An example of the application of the Fisher and Yates method, for an
interval that ranges from 1 to 5, is illustrated in Table 4.1.
Figure 4.3: Schematic representation of the possible scenario generation.
Table 4.1: Example of the Fisher and Yates shuffle algorithm. The permuted
interval ranges from 1 to 5. The red cells correspond to the number that is
selected in each iteration.
Step Random Number Sequence Result

1 4 1 2 3 4 5 4
2 1 1 2 3 4 5 41
3 3 1 2 3 4 5 415
4 2 1 2 3 4 5 4153
4 1 1 2 3 4 5 41532
43
The generated seismic lithofacies indicator 1D array is the template used
for the rock property 1D arrays generation. The property 1D arrays are produced
in an order that starts with the lithology and porosity, and finishes with the elastic
properties. Each cell in the property 1D array is filled with the property
information calculated from the multivariate PDF and defined by the Markov
Chain. The accuracy of the rock properties model relies on the idea that the
multivariable PDF truly represents the statistics of the target lithofacies.
We have also included some additional constraints for our inversion:
a) Fluid contacts: For geological models with local fluid contacts, all samples
of gas reservoirs should be above the oil samples and all samples with oil
should be above the water samples.
b) Geological age: The seismic lithofacies geometry can be constrained in a

given order that follows a stratigraphic column.
c) Layer thickness: The thickness of the layers can be constrained to any

given range.
4.4. Synthetic Seismic Generation
The probabilistic inverse problem defined by Equation 1.1 estimates the a
posteriori PDF ( ( )) from the a priori PDF of the data ( ( )) and the model
( ( )). Equation 1.1 is valid when the information in the space of measured
parameters (data) has been obtained independently of the prior information in the
model space (Tarantola, 2005). In section 1.1, we introduced forward modeling
operator ( ( )) that predicts data (Equation 1.2) from the model. The predicted
44
values are not necessary identical to the observations since there are
uncertainties related to the measurements and the proposed model.
In the present study the forward modeling operator is implemented as a
convolutional model for the synthetic trace generation. This model is expressed
as follows:
( ) ( ) ( ), (4.1)
where:
( )= Synthetic seismic trace,
( )= Angle dependent reflection coefficients in time domain,
( )= Angle dependent extracted wavelet, and
‘*’ = Convolution operator.
The convolutional model of the seismic trace assumes that the latter is the
superposition of the reflection coefficients filtered by the bandwidth of the
recorded wavelet (Sheriff, 1989). Since the elastic property 1D arrays are
generated in the depth domain, they have to be stretched to two-way time using
a depth-to-time function. We assumed here that the depth-to-time conversion
function is part of the input data. Another important part of the input data is the
set of seismic wavelets. Seismic wavelets have to be previously extracted from
the measured seismic data at the same angle ranges at which the synthetic data
have been generated.
45
After the elastic properties arrays are converted to two-way travel time,
they are interpolated to the same regular sampling as the input seismic data.
This is a very important process since the time sampling acts as a frequency
operator that defines the maximum vertical resolution of the inverted earth model
(see Figure 4.4). Finally, a Nyquist filter is applied to the time-stretched elastic
models to avoid any possible aliasing in the time domain.
Figure 4.4: Comparison among the depth, time and regularized time 1-D P-wave
velocity arrays. The black circles show how the regularization in the time domain
can affect the elastic properties of the modeled subsurface.
Once the elastic properties arrays in time have been generated, the angle-
dependent reflectivity profiles are produced using the Aki and Richards linear
approximation (Aki and Richards, 1980) of the Zoeppritz equation. The Aki and
46
Richards linear approximation can be expressed in terms of P-wave velocity, S-
wave velocity and bulk density (ρ) as follows:
( ) ( ) [ ( )] ( ) (4.1)
where:
ΔVp=Vp2-Vp1 Vp=0.5(Vp1+Vp2)
ΔVs=Vs2-Vs1 Vs=0.5(Vs1+Vs2)
Δρ=ρ2-ρ1 ρ=0.5(ρ1+ρ2)
ϴ=0.5(ϴ1+ ϴ2);
Vp, Vs, ρ and ϴ are the P-wave velocity, S-wave velocity, bulk density and
incidence angle, respectively. The sub-indexes 1 and 2 refer to the upper and
lower layers correspondingly. Equation 4.1 can be rewritten as:
( ) ( ) (4.2)
where A, B and C are the AVO intercept, AVO gradient and AVO curvature. We
used the first two terms of AVO Equation 4.2 for the estimation of the angle-
dependent reflectivity arrays. The two-term Aki and Richards AVO equation is
valid for a maximum incident angle of about 36°.
Finally, the synthetic trace is calculated by the convolution of angle-dependent
reflectivity converted to two-way time and the angle-dependent wavelet. Figure
4.5 shows a schematic representation of the synthetic traces generation step.
47
Figure 4.5: Schematic representation of the synthetic traces generation
algorithm.
4.5. The Misfit Function
The final step in the 1D stochastic inversion scheme is the estimation of
the misfit function between the real and synthetic seismic traces. The synthetic
trace is compared to the real seismic trace and if the difference between them is
less than a given threshold the model is accepted; otherwise, the model is
erased and a new one is generated.
In the proposed algorithm, the similarity analysis is performed in a
cascade sequence that starts with the nearest and finished, with the farthest
angle-stacked traces. Once a trace is accepted the algorithm analyzes the next
angle trace until all the angle-dependent reflectivity traces are accepted. There
are many criteria to estimate the degree of similarity between the synthetic and
48
real traces; we used L2-norm to measure the similarity between the real and
synthetic traces. The L2-norm can be expressed as follows:
√∑ ( ) , (4.2)
where: Ti Syn= Synthetic trace, Ti Real= Real trace, and i= ith Element of the
1-D trace array.
An example of one trace misfit analysis is shown in Figure 4.6. The top
graph shows the real (black) and synthetic (blue) traces. The difference between
the synthetic and real traces is displayed in the center graph. Finally, the
cumulative L2-norm is plotted in the bottom graph. In our proposed algorithm, we
select the models that minimize the L2-norm.
Figure 4.6: Misfit calculation example. In this synthetic example the cumulative
2
L -norm shows the intervals with larger misfit.
49
The probable models generated by the proposed inversion scheme are a
valid solution to the studied inverse problem; all of them honor the input rock
physics model as well as the measured seismic data.
50
5. SYNTHETIC EXAMPLES
5.1. Introduction
The central problem in seismic reservoir characterization is the prediction
of lithology and saturation from seismic and well log data. In this chapter we
present the results of the application of our proposed inversion scheme on two
different sets of synthetic data. The synthetic data sets were generated based on
real well data with the goal of testing the inversion in two different geological
environments. The first synthetic data set is composed of low porosity high
impedance sand. The goal of the second example is to replicate a high porosity
low impedance clastic environment. In this example, the target reservoir rock is
characterized by weakly compacted uncemented sandstones.
In both synthetic tests, the model data space is represented by the
multivariate PDFs that were generated from the synthetic log and extended by
theoretical models. For instance, changes in pore fluid properties were modeled
using Gassmann fluid replacement theory. The rock’s properties in both synthetic
log data sets were based on real measured data. Next, the data space was
generated using synthetic seismic data modeled from synthetic elastic well logs.
As discussed in Section 4.4, the forward modeling operator was based on the
convolutional model of the seismic trace.
51
The content of this chapter was partially presented in the paper IPTC-18040 of the 2014 International Petroleum
Technology Conference in Kuala Lumpur.
5.2. Synthetic Test 1: Cemented Sand
The first synthetic data set was based on real well log data measured in a
Cretaceous interval; the in situ well logs are displayed in Figure 5.1. The
synthetic model was formed by two sands encapsulated into a shale sequence.
The top upper sand is located at 3,347 meters depth with a thickness of 40
meters. The lower sand is thinner with a thickness of 20 meters and located just
32 meter bellow the base of the upper sand. The target reservoirs are cemented
sand with average porosity of 16%. The background P-wave impedance is about
10 km*kg/sec*cc. The average P-impedance corroborates that the sands are well
compacted and cemented.
Figure 5.1: Model 1 synthetic well logs showing (from left to right) facies
indicator, shale volume, P-wave velocity, bulk density, S-wave velocity, porosity,
P-wave impedance and S-wave impedance.
52
In this kind of elastic scenario, fluid variations have very small influence on
the elastic properties of the target reservoirs. In high-impedance reservoir
settings, the goal of quantitative interpretation is usually the prediction of gross
reservoir thickness and reservoir quality. In our study here, the gross reservoir
thickness is expressed by the fraction of the potential reservoir rock in the
inverted interval. The reservoir quality is defined by two properties, which are the
fraction of shale (shale volume) and the total porosity of the sandstone. The
study reservoir properties are described by non-parametric histograms and
descriptive statistical estimators.
The elastic fluid properties profiles for brine, oil and gas were estimated
using the Batzle and Wang (1992) equations and assuming pressure and
temperature gradients of 9.72 MPa/km and 22.3 °C/km. Oil gravity, gas gravity
and salinity were 36° API, 0.6 and 8,000 ppm respectively. Figure 5.2 shows the
P-wave and density profiles for three possible pore fluids. Once the fluid
properties are estimated, the fluid replacement is performed using the Gassmann
relations. The fluid substitution was performed from brine to 90% hydrocarbon
saturated sands.
53
Figure 5.2: Synthetic model 1 pore fluids velocity and density profiles.
The final seismic lithofacies multivariate PDF projections for the Vp-Vs,
Vp-Rhob, and PhiT-Vp spaces are shown in Figure 5.3. The effect of the fluid
variations on the elastic properties of the permeable rocks is negligible. However,
there was good lithology discrimination in all projected spaces. Hence, the test
was planned focus on the prediction of gross reservoir thickness and reservoir
quality.
54
(Upper), Vp-Rhob (Center), and PhiT-Vp (Lower) spaces. Green, blue, brown
and red contours denote shales, brine sands, oil sands and gas sands,
correspondingly. Some empirical models are plotted for comparison including
Castagna’s mudrock line, Castagna’s 1993, Han’s 1986, Gardner’s sand,
Gardner’s shale, Voigt bound, Reuss bound and Nu’s modified Voigt bound for
quartz mixture.
The angle-dependent reflectivity curves were generated from the synthetic
logs converted to time. They showed strong positive reflection coefficients at the
top of both sands. It was clearly evident that one of the challenges in this test
was to accurately predict the reservoir geometry in the target interval. Figure 5.4
55
shows the model 1 angle dependent reflectivity profiles. It can be observed how
the reflectivity of the top and base of the sands decrease with the angle of
incidence. This is typical reflectivity response of high impedance sands.
Figure 5.4: Model 1 angle dependent reflectivity profiles.
Next, we generated the synthetic angle stack traces by convolving the
angle-dependent reflectivity with a Ricker wavelet with a central frequency of 25
Hz (Figure 5.5).
56
Figure 5.5: Ricker wavelet with a central frequency of 25 Hz.
For this test we assumed the stationarity of the wavelet, meaning that
wavelet did not vary with incident angle or propagation time. The results of this
forward modeling are shown in Figure 5.6. A set of five angle traces were
generated equally spaced in an interval that ranged from 0° to 32°. This angle
gather was used as the measured seismic data for this synthetic test.
Figure 5.6: Synthetic angle gather generated from the synthetic model 1.
57
Once the “real traces” were generated the next step was to produce the
multivariate PDF for each of the analyzed lithofacies. The 2D projections of the
multivariate PDF (Figure 5.6) show that the sands have larger p-wave velocity,
lower Vp/Vs ratio and larger density than the surrounding shales. The
observations in Figure 5.7 indicate that there is an excellent lithofacies
discrimination based on the elastic properties of the analyzed statistical
populations.

sands (left) and shales (right) calculated from the synthetic model 1.
58
The inversion was performed for one million realizations using the model
shown in Figure 5.6 as the measured data and the multivariate PDFs shown in
figure 5.7 as the rock physics models. Modeled traces with similarity factor (α)
smaller than 0.9 were selected as possible solutions for the stochastic inversion
of the synthetic data. The comparison between the synthetic seismic data and
the 1,112 possible solutions revealed the close similarity between the synthetic
measured data and the probable models seismic traces. The residuals were
small, the time arrivals (Figures 5.8) and waveforms (Figures 5.9) similar.
2
comparison. The selected traces have L -norms smaller than 0.9.
59
models with α<0.9.
The seismic lithofacies possible models with α<0.9 are displayed in Figure
5.10. It can be observed that most of the sand samples are located inside the
limits of the upper and lower sands. In fact, these sands are well-delimited by
most of the probable inversion solutions. As discussed Section 5.1 we are
interested in the prediction of gross reservoir thickness and reservoir quality.
Gross reservoir thickness is defined by the fraction of sand in the target interval.
Figure 5.11 shows the PDF of the sand fraction for all probable models with
60
α<0.9. The most expecting value of the sand fraction is 0.13, which is the same
fraction defined in the synthetic model. However, there is not a sole solution,
since the inversion produces a range of likely solutions that are constrained by
the rock-physics model.
Figure 5.10: Probable models with α<0.9. The black, red, yellow and green lines
of lower sand, respectively.
61
Figure 5.11: Sandstone fraction for probable models with α< 0.9.
Once the geometrical distribution of the two seismic lithofacies was
calculated from the probable models, we were able to define the key rock
properties for each lithofacies, and especially for the sandstones. For this
particular case, total porosity was a very important property for the definition of
reservoir quality. From the total-porosity models in Figure 5.12 we were able to
calculate some descriptive statistical coefficients; for this particular test, the most
likely sand total porosity was about 0.15, with a standard deviation of 0.7. From
the total porosity and volume of shale probable models (Figure 5.13), it was
possible to estimate the effective porosity of the target reservoir (Figure 5.14).
Finally, the elastic property solutions are shown in Figure 5.15.
62
Figure 5.12: Total porosity of the probable models with α<0.9. The black, red,
of lower sand and base of lower sand, respectively.
Figure 5.13: Volume of shale probable models with α<0.9. The black, red, yellow
and green lines represent the top of upper sand, base of upper sand, top of lower
sand and base of lower sand, respectively.
63
Figure 5.14: Effective porosity probable models with α<0.9. The black, red,
of lower sand and base of lower sand, respectively.
Figure 5.15: Elastic models with α<0.9. The black, red, yellow and green lines
of lower sand respectively.
64
As has been mentioned before, all the solutions generated by the
proposed seismic inversion scheme are probable answers to the analyzed
seismic inverse problem. Therefore, the statistics of the solutions have more
significant relevance than the models themselves. The comparison between the
rock properties of the synthetic model and the solutions (Figure 5.16) shows a
very close similarity between the synthetic data and the inversion solutions. The
synthetic model used in this exercise is realistic and it is based on measured well
log information.
Figure 5.16: Normalized histograms comparison between the rock properties of

the synthetic model and the probable solutions, for sand’s shale volume (upper
left), sand’s total porosity (lower left), shale’s shale volume (upper right) and
shale’s total porosity (lower right). Red and blue bars correspond to synthetic
model and solutions, respectively.
65
It is clear that if α decreases the similarity of the possible models will
increase. However, the number of possible models will decline as well. Figure
5.17 shows the possible solution models for α<0.045; in this particular case, we
obtained 78 possible models instead of the 1,112 calculated with α<0.9. This is
an important characteristic of the proposed inversion scheme; this is further
discussed in Section 7.4.
Figure 5.17: Probable models with α<0.045. The black, red, yellow and green
lines represent the top of upper sand, base of upper sand, top of lower sand and
base of lower sand, respectively.
5.3. Synthetic Test 2: High Porosity Sands
The first synthetic test proved the viability of the application of the
proposed inversion scheme to predict the geometry and quality of the reservoir of
66
the target formation. The goal of the second synthetic test is to replicate a high-
porosity, low-impedance clastic environment. In this example, the target reservoir
rock was characterized by uncemented sandstone with an average porosity of
31%. In this geological setting, fluid properties variations have very strong effect
on the elastic properties of the studied sands. The associated hydrocarbon
seismic signatures of these high-porosity reservoirs are usually anomalous high
amplitude reflections called “bright spots”. In this test, the objective was to define
the fraction of sand, the fraction of hydrocarbon sand, the fluid contact, the
volume of shale and the effective porosity of the analyzed sand.
The synthetic well logs, displayed in Figure 5.18, show the in situ case in
black line, brine case in blue and the gas case in red. The fluid replacement was
performed using the Gassmann relations. The elastic fluid properties profiles for
brine, oil and gas (Figure 5.19) were estimated using the Batzle and Wang
(1992) equations and assumed pressure and temperature gradients of 9.72
MPa/Km and 20 °C/Km. The gas gravity and salinity were 0.6 and 10,000 ppm
respectively. The well logs indicated that the brine-saturated sandstone had
larger impedance than the surrounding shale. However, when the pore fluid was
substituted from brine to a mixture of 90% gas and 10% brine, the sandstone’s
impedance became lesser than the shale’s acoustic impedance. Again, the rock-
physics diagnostic plot (Figure 5.20) shows the most important rock properties
for the three analyzed seismic lithofacies; in general, the elastic discrimination
among the seismic groups was good.
67
1950
2000
2050
2100
2150
Depth [m]
2200
2250
2300
2350
2400
2450
0 0.5 1 0 0.2 0.4 1.5 2 2.5 1.5 2 2.5 0.5 1 1.5
Vshale [fraction] Porosity [fraction] Vp [km/sec] Rhob [g/cc] Vs [km/sec]
Figure 5.18: Model 2 synthetic well logs showing (from left to right) facie
indicator, shale volume, porosity, P-wave velocity, bulk density and S-wave
velocity. Black, blue and red lines correspond to in situ, brine and gas
saturations.
1950
2000
2050
2100
2150
DEPTH [m]
2200
2250
2300
2350
2400
2450
0.5 1 1.5 2 0 0.5 1 1.5
Vp [km/sec] Rhob [g/cc]
Figure 5.19: Synthetic model 2 pore fluids velocity (left) and density profiles
(right). Blue and red lines correspond to brine and gas, respectively.
68
(Upper), Vp-Rhob (Center), and PhiT-Vp (Lower) spaces. Grey, blue, and red
contours denote shales, brine sands and gas sands, correspondingly. Some
empirical models are plotted for comparison, including Castagna’s Mudrock line,
Castagna’s 1993, Han’s 1986, Gardner’s sand, Gardner’s shale, Voigt bound,
Reuss bound and Nu’s modified Voigt bound for quartz mixture.
The angle-dependent reflectivity plot (Figure 5.21) shows the main
impedance contrasts in the studied interval; the top of the gas sand is delineated
by a negative reflection coefficient with strong negative gradient. The largest
positive reflection coefficient corresponds to the gas-brine fluid contact within the
sandstone. This seismic interface has a relative small positive AVO gradient.
69
Finally, the base of the brine-saturated sand is defined by a negative reflectivity
with a strong negative gradient.
Figure 5.21: Model 2 angle-dependent reflectivity profiles.
The synthetic angle gather (Figure 5.22) was generated by convolving a
25 Hz Ricker wavelet (Figure 5.5) with the reflectivity logs displayed in Figure
5.22. As was expected, the most important reflection is the gas-brine fluid contact
situated at approximately 2.200 seconds. Next, the multivariate probability
functions were generated for the shale, the brine sand and the gas sandstone
seismic lithofacies (Figure 5.23).
70
Figure 5.22: Synthetic angle gather generated from synthetic model 2.

gas sands (left), shales (center) and brine sands (right), calculated from synthetic
model 2.
71
The inversion was five million realizations with the constraint that all brine
samples should be above the deeper gas sample. This is important because we
assumed a unique fluid contact in the whole interval; the models with L2-norms
smaller than 0.88 were selected as possible solutions for the inversion. The
comparison between the synthetic model and the 664 obtained solutions (Figure
5.24) shows a close similarity between the synthetic input data and the inversion
results. The waveform comparison for four inverted solutions and the synthetic
model (Figure 5.25) demonstrate that the main events are well reproduced by the
inversion algorithm.
0.08
2
0.06
2.05
2.1 0.04
Amplitude [Dimensionless]
2.15 0.02
Time [sec]
2.2 0
2.25 -0.02
2.3 -0.04
2.35 -0.06
2.4 -0.08
100 200 300 400 500 600

Number of Traces
2
comparison. The selected traces have L -norms smaller than 0.88.
72
models with α<0.88.
The seismic lithofacies distributions for the 664 solutions are shown in
Figure 5.26. A visual inspection of the seismic lithofacies distributions shows a
large consistency for the depth of the top of the sand, fluid contact and base of
the sand. We generated histograms of the depth of the three seismic lithofacies
interfaces to evaluate the accuracy of the proposed inversion scheme (Figure
5.27).
Table 5.1 shows the comparison between the synthetic data and the
depths obtained by the inversion; these results demonstrate that the inversion
scheme is very accurate.
73
Figure 5.26: Probable models with α<0.88. The green, black and white represent
the shales, gas sands and brine sandstone seismic lithofacies, respectively.
400 400 350
350 350
300
300 300
250
250 250
200
Counts
Counts
Counts
200 200
150
150 150
100
100 100
50
50 50
0 0 0
2155 2160 2165 2170 2175 2190 2195 2200 2205 2210 2210 2220 2230 2240
Depth of Top of the Sand [m] Depth of Gas-Water Contact [m] Depth of Base of the Sand [m]
Figure 5.27: Depth of top of the sand (left), gas-water contact (center) and base
of the sand (right) histograms.
74
Table 5.1: Comparison between the synthetic data and the depths obtained by
the inversion.
Gas-Water
Top of Sand Base of Sand
Contact
Synthetic Model
2166.00 2202.00 2222.00
Value [m]
Mean [m] 2165.80 2198.70 2222.30
Median [m] 2166.00 2198.00 2222.00
Standard Deviation 2.82 2.63 3.21
Variance 7.97 6.92 10.34
Figures 5.28 to 5.31 show total porosity, shale volume, and effective
porosity, respectively. The effective porosity was calculated from the total
porosity and volume of shale properties. The elastic property solutions are shown
in figure 5.32. The fluid contact is easily observed in all the elastic models.
Total Porosity Vertical Models
2000 0.32
2050 0.3
Total Porosity [fraction]

2100 0.28
2150 0.26
Depth [m]
2200 0.24
2250 0.22
2300
0.2
2350
0.18
2400
0.16
100 200 300 400 500 600
Number of Traces
Figure 5.28: Total porosity of the probable models with α<0.88.
75
Volume of Shale Vertical Models
0.9
2000
0.8
2050
0.7
Shale Volume [fraction]

2100
0.6
2150
Depth [m]
0.5
2200
0.4
2250
0.3
2300
0.2
2350
2400 0.1
100 200 300 400 500 600

Number of Traces
Figure 5.29: Volume of shale probable models with α<0.88.
Effective Porosity Vertical Models
0.3
2000
2050 0.25
Effective Porosity [fraction]

2100
0.2
2150
Depth [m]
2200 0.15
2250
0.1
2300
2350 0.05
2400
0
100 200 300 400 500 600
Number of Traces
Figure 5.30: Effective porosity probable models with α<0.88 calculated from the
total porosity and shale volume properties.
76
Figure 5.31: Elastic models with α<0.88.
5.4. Conclusions
The synthetic test results presented in this chapter demonstrate the
feasibility of the estimation of rock properties by our proposed probabilistic
inversion approach. The first example showed the capability of the novel
77
inversion scheme to estimate two critical rock properties for hydrocarbon
exploration: porosity and reservoir fraction.
The second synthetic test included the fluid prediction as an additional
rock property to be estimated by the inversion. The outcome demonstrated that
the gas-filled sandstone can be easily discriminated from the shales and brine-
saturated seismic lithofacies. We also showed the accuracy of the method in the
estimation of the depth of the seismic lithofacies interfaces.
In conclusion, the results of these synthetic tests proved the feasibility of
using the proposed stochastic inversion technique to calculate the properties of
seismic lithofacies based on multivariate probabilistic rock physics models,
reflection seismic data, using Markov Chain Monte Carlos Simulation (MCMCS)
to sample the model space.
78
6. 1D REAL DATA APPLICATION
6.1. Introduction
This chapter presents the application of our proposed seismic stochastic
inversion algorithm to a real data set. The data set includes a group of well logs
and a 2D line of prestack seismic data extracted from a 3D volume crossing the
exploratory well. The studied stratigraphic sequence corresponds to tertiary
marine sediments deposited on a platform margin. From well-log data and the
geological information, we identified two main lithologies: sandstones and shales.
The potential reservoirs are unconsolidated sandstones with total porosity that
ranges from 35% to 39%.The pore space can be filled by gas, low saturation gas
(LSGS) or brine. Each combination of fluid and lithology is considered an
independent seismic lithofacies. Therefore, the target interval is made up of four
seismic lithofacies: shales, gas-filled sands, low saturation gas sands (LSGS)
and brine sands. Figure 6.1 shows the in situ saturation logs for the two hundred
and twenty meter study interval. The target interval has two sands: the upper
sand is interceded with total thickness of twenty meters and a sand fraction of
73%; the lower reservoir is eighty meters of blocky sand with three fluid faces:
gas, LSGS and brine.
The seismic data included raw PSTM gathers with a time sampling of 2
msec. The near angle stack generated from 0° to 15° is shown in Figure 6.2; the
top of the strong seismic anomaly related to the top of the gas reservoir is
79
represented by large negative seismic amplitude. The top of the upper and lower
sands are located at 72 ms and 104 ms, respectively. The goal of this inversion
was to predict total porosity, volume of shale, effective porosity, and fluid type
distribution in the study interval.
Figure 6.1: Well 1 logs for in situ saturation.
Figure 6.2: Near angle stack; the top of the upper and lower sands are located at
72 ms and 104 ms, respectively.
80
6.2. Seismic Data Conditioning
The prestack seismic data were conditioned to remove residual coherent
noise and to improve the signal-to-noise ratio for the stochastic seismic inversion.
This was an important stage, since the forward modeling operator assumed that
all the reflections corresponded to primary reflections and random noise, and no
coherent noise was present in the seismic data. The seismic data was also
resampled to one millisecond to match the time sampling of the synthetic models.
The angle gather at Well-1 location is displayed in Figure 6.3; the amplitude
extraction at the top of the upper sand shows a strong AVO effect associated
with the top of the gas sand.
Figure 6.3: Angle gather at the Well-1 location (Upper). The amplitude extraction
at the top of the upper sand shows a strong AVO effect (Lower).
81
After the seismic data was conditioned and converted to angle gathers,
the well logs were converted to time and the well-to-seismic tie was performed
(Figure 6.4). The well log tie shows close similarity between the zero offset trace
and the acoustic synthetic trace (right track). The angle-dependent extracted
wavelets are shown in Figure 6.5. The L2-norm calculated from the real and
synthetic traces is α=0.88. Therefore, we assumed that any inversion solution
with α≤0.88 was a possible solution to the studied real data application.
Figure 6.4: Well logs converted to time showing the well to seismic tie (right
track); the black line corresponds to the measured seismic trace and blue the
synthetic seismogram.
82
1
oo
0.8
8o
16o
0.6
24o
32o
0.4
0.2
-0.2
-0.4
-100 -80 -60 -40 -20 0 20 40 60 80 100
Time [ms]
Figure 6.5: Angle-dependent extracted wavelets.
6.3. The Model Space Generation
The available well-log information was used to generate the multivariate
PDF for the four observed seismic lithofacies: shale, gas sand, low gas
saturation sand and brine sand. Fluid replacement based on Gassmann’s
relation was performed to better understand the effect of fluid variations in the
target reservoir sands. Figure 6.6 shows the well logs for the in situ, 90% gas,
and 100% brine-saturated sands.
The 2D projections of the multivariate PDF (Figure 6.7) show the feasibility
of discriminating most of the seismic lithofacies groups based on their elastic
properties. Only the low saturation gas sands cannot be easily discriminated from
gas sands.
83
Figure 6.6: Well 1 logs after fluid substitution. Black, red and blue lines
correspond to in situ, 90% gas and 100% brine-saturated sands.
Shale Brine Sand Gas Sand LSGS

1.4 1.4 1.4 1.4
Vs [km/sec]
Vs [km/sec]
Vs [km/sec]
Vs [km/sec]
1.2 1.2 1.2 1.2
1 1 1 1
2 2.5 3 2 2.5 3 2 2.5 3 2 2.5 3
Vp [km/sec] Vp [km/sec] Vp [km/sec] Vp [km/sec]
2.2 2.2 2.2 2.2

Rhob [g/cc]
Rhob [g/cc]
Rhob [g/cc]
Rhob [g/cc]
2 2 2 2
1.8 1.8 1.8 1.8
2 2.5 3 2 2.5 3 2 2.5 3 2 2.5 3

Vp [km/sec] Vp [km/sec] Vp [km/sec] Vp [km/sec]
3 3 3 3
Vp [km/sec]
Vp [km/sec]
Vp [km/sec]
Vp [km/sec]
2.5 2.5 2.5 2.5
2 2 2 2
0.2 0.3 0.4 0.2 0.3 0.4 0.2 0.3 0.4 0.2 0.3 0.4
PhiT [fraction] PhiT [fraction] PhiT [fraction] PhiT [fraction]
Figure 6.7: 2D Projections of the multivariate probability distribution function

generated from Well 1 log data and Gassmann fluid substitution.
84
6.4. The Seismic Stochastic Inversion
The rock property models were generated assuming that all gas samples
were located above the low saturation gas sands and all LSGS were situated
above the brine sand. The fluid contacts can be situated at any depth as long as
the order of the fluid interfaces is followed. The properties of the seismic
lithofacies were assigned from the multivariate PDFs by using the MCMCS
engine to sample them. The proposed inversion algorithm generated data in two
domains; the first is the rock physics simulation domain that is generated in
depth. The second was the forward-modeling domain that was built in two-way
travel time. For this real data application, the vertical sampling of the rock physics
simulation grid (depth) was fixed to one meter and the two-ways travel time
sampling for the forward model grid was set to one millisecond. The depth and
time grids were selected to minimize the resampling effect on the well-log data.
Five million realizations were performed and 320 models with normalized
α≤0.88 were selected. The comparison between the real traces and the 320
solutions for 0° and 32° incident angles are shown in Figure 6.8.
85
Angle Traces at 0o
2
50 1
Time [msec]
100 0
150 -1
200 -2
50 100 150 200 250 300
Number of Traces
Angle Traces at 32o
2
50 1
Time [msec]
100 0
150 -1
200 -2
50 100 150 200 250 300
Number of Traces
Figure 6.8: Comparison between real angle traces (upper and lower left) and the
320 solutions for 0° (upper right) and 32° (lower right) incident angles.
Figure 6.9 shows that the similarity among the traces is large enough to
assume that all of 320 model traces were generated from probable geological
models. The upper and lower sands are well-delimited by most of the probable
inversion solutions. However, the gas-LSGS fluid contact can be located at any
depth because of the impossibility of elastically discriminating between gas and
low saturation gas sands. The possible models plot shows that all the possible
solutions for the LSGS-brine sand fluid contact have the fluid interface located in
the lower sand. In fact, most solutions have the fluid contact situated more than
20 meters below the top of the lower sand.
86
Figure 6.9: Probable models with α≤0.88. Grey, black and white represent
shales, gas sands and low saturation gas sands, respectively.
We are interested in the prediction of total porosity, shale volume, and
effective porosity prediction from the seismic inversion, as has been mentioned.
Figures 6.10, 6.11 and 6.12 show the total porosity, volume of shale, and
effective porosity for the 320 solutions. The elastic models shown in Figure 6.13
demonstrate that there are almost no differences in the elastic properties of the
gas sands and low saturation gas sands. In fact, the main differences are found
in the density rather than in the elastic moduli.
87
Figure 6.10: Total porosity of the 320 probable models.
Figure 6.11: Volume of shale of the 320 probable models.
88
Figure 6.12: Effective porosity of the 320 probable models.
Figure 6.13: Possible elastic models with α≤0.88.
89
6.5. Analysis
The combination of the probable models for the target zone was used for
the analysis of rock properties. Figure 6.14 shows the probability of the sand plot;
larger values mean higher possibility of having a sand sample at a given depth. It
can be observed that the upper sand has less probability than the lower sand,
due to its sand/shale interbedded geometry. We can assume any probability
cutoff to define the most probable sand/shale geometry; however, it is important
to mention that the larger the cutoff the smaller the sand fraction in the analyzed
interval; we chose a conservative value of 0.2 as the threshold for this real data
set. The comparison between the real and the estimated top values for
probability equal to or larger than 0.2 is shown in table 6.1. The comparison
confirms the accuracy of our proposed inversion scheme for lithology prediction.
90
20
40
60
80
100
Depth [m]
120
140
160
180
200
220
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Probability of Sand
Figure 6.14: Probability of sand vertical plot.
Table 6.1: Comparison between the real and the estimated lithology top values
for probability larger than 0.2.
Inverted Tops [m] Real Tops [m]

Top Upper Sand 80 79
Base of Upper
103 98
Sand
Top Lower Sand 116 114
Base of Lower
176 181
Sand
91
Based on the inversion solutions for total porosity and effective porosity,
we generated the most probable vertical plot for both rock properties (Figure
6.15). The comparison shows that the inverted porosities underestimated the
larger porosity values and overestimated the lower observed well log porosity.
This smoothing effect is stronger in the upper sand where extreme property
values are further apart.
0 0
20 20
40 40
60 60
80 80
100 100
Depth [m]
Depth [m]
120 120
140 140
160 160
180 180
200 200
220 220
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Total Porosity [fraction] Effective Porosity [fraction]
Figure 6.15: Total porosity and effective porosity comparison. The black lines
correspond to the well log information and the blue lines to the inverted rock
properties.
92
The most-likely fluid geometry (Figure 6.16) was generated from the 320
probable models section (Figure 6.9) and constrained by the most expected sand
profile (Figure 6.14). The comparison between the real fluid interfaces and the
inverted fluid contact depths (Table 6.2) shows a surprisingly close agreement
with the well data.
Figure 6.16: Most expected fluid seismic lithofacies (left) and probability of fluid
type (right).
Table 6.2: Comparison between the real and the estimated fluid contact depths.
Inverted Contacts [m] Real Contacts [m]

Gas - LSGS Contact 143 147
LSGS - Brine 168 176
93
Finally, the elastic properties probable models generated from the
stochastic inversion are compared with the Well-1 elastic logs in Figure 6.17. The
results of the inversion are remarkable similar to the well log measured data.
However, the S-wave velocity of the interbedded shales within the upper sand
section was not well predicted. These shales have a very low Vp/Vs ratio
compared to the rest of the shales.
0 0 0
20 20 20
40 40 40
60 60 60
80 80 80
100 100 100

Depth [m]
Depth [m]
Depth [m]
120 120 120
140 140 140
160 160 160
180 180 180
200 200 200
220 220 220

1.5 2 2.5 3 3.5 1.5 2 2.5 1 1.2 1.4 1.6 1.8
Vp [km/sec] Rhob [g/cc] Vs [km/sec]
Figure 6.17: Comparison between the most expected elastic models (blue line)
against well logs (black lines).
94
6.6. Conclusions
The outcomes of the 1D real data application of the proposed stochastic
inversion are remarkable similar to the available well log measured data. It is
important to emphasize that Well-1 was not used to constrain the inversion; it
was only used to populate the model space. The mean of the inverted rock
properties tends to be smoother than the measured values; in fact, the most
expected inverted properties have a tendency to underestimate larger rock
property values and overestimate lower property measurements. This
‘smoothing’ effect is stronger in layers where extreme property values are further
apart. This effect is observed in all the estimated rock properties.
The S-wave velocity of the shale member of the interbedded upper sand
was underestimated by the inversion; this suggests that an additional seismic
lithofacies might be included in the model space. Finally, the proposed inversion
approach demonstrates the viability of the integration seismic data, well log
information, and rock physics models in a practical workflow to estimate rock
properties for hydrocarbon exploration or development.
95
7. 3D REAL DATA APPLICATION
7.1. Introduction
In this chapter, we introduce the first 3D application of a real data set
using the proposed inversion algorithm. The real data seismic inversion is
performed using the same data set presented in the Chapter 6. The main
difference between the 1D and 3D stochastic inversion algorithms is the use of
geostatistics and well log data to constrain the outcomes.
In the first part of the chapter, we introduce some basic concepts of
geostatistics that are used to create lateral continuity to the stochastic seismic
inversion outcomes. The second part of the chapter shows the application of the
proposed 3D stochastic inversion algorithm to the 2D line of prestack seismic
data extracted from a 3D volume and supported by the Well-1 information.
Finally, we discuss the convergence criteria and the efficiency of the proposed
inversion algorithm.
7.2. The Kriging Estimator
Geostatistics is a set of mathematical tools used to describe the lateral
continuity of natural phenomena. Geostatistics was originally developed to
estimate probability distributions of mineral concentrations for mining
prospecting. Classical geostatistics is strongly related to the Kriging interpolation
algorithm (Isaaks and Srivastava, 1989). The Kriging algorithm is a family of
96
linear unbiased estimators that predict the value of a property in any given
location by computing a weighted average of the observed values of the studied
properties and their spatial distributions. All the Kriging algorithms assume that
the observed data are the result of a random process; then the objective of the
Kriging estimator is to minimize the error variance ( ).
The Kriging estimators are defined by the treatment of the trend
component (mean). In this particular application, we used the Ordinary Kriging to
include lateral continuity to the stochastic inversion. Ordinary Kriging assumes
that the local mean may vary greatly over the studied area. In this case, the
system of equations is:
∑ ( ) ( ) ( ) ( )
{ , (7.1)
∑ ( ) ( )
where:
uα, uβ = Location vectors for known location,
u = Location of the unknown location,
h = uα-uβ,
CR(h) = Covariance function,
λα(u) = Kriging weight.
In general, the weight of each data point decreases with increasing
distance to the location that is being estimated. The weights λα are calculated
from the covariance function that is generally derived from the semivariogram
97
model. The semivariogram model is a function that describes the degree of
spatial dependence of the target property. It is defined as the variance of the
difference between known values at two locations across the target area. The
Kriging estimator needs valid semivariogram models to compute the covariance
at any distance h. In isotropic environments, the correlation between the spatial
process at two point pairs p1 and p2 only depends on the separation distance h
and not on the orientation of the two points. However, real data often show
directional effects. Particularly in geologic applications, measurements along a
particular direction may be highly correlated, while the perpendicular direction
shows little or no correlation. This phenomenon is known as anisotropy.
There are two types of anisotropy. The simplest type occurs when the
covariance form is the same in all directions but the range S changes. In this
case, there is a single sill, but the semivariogram reaches the sill in a shorter lag
distance along a certain direction. In anisotropic settings h can be estimated
using the following relation:
( ), (7.2)
where R is the ratio of between the larger r1and the lower seal r2:
. (7.3)
In this real data application we use a spherical semivariogram to estimate
the Ordinary Kriging weights. The spherical semivariograms are defined as:
98
( ) ( ) ( ) , (7.4)
where h is the distance vector and a is a fitting constant defined from the
observed data.
7.3. 3D Real Data Application
As has been mentioned before, the 3D application integrates the 1-D
stochastic inversion algorithm with a Kriging engine to produce the lateral
constraint rock properties model. Figure 7.1 shows the simplified stochastic
inversion algorithm; the inversion scheme includes an additional choice to select
between the Markov Chain simulation for rock properties or the geostatistical
estimation of rock properties. In this real test we inverted the 50 traces that
constitute the 2D seismic line (Figure 6.2) using the Well-1 as initial well data and
the statistical rock physics model shown. To estimate the covariance we used an
anisotropic spherical semivariogram with a 0.1 nugget, a sill of 0.9 and ranges of
500 meters and 50 meters in horizontal and vertical directions respectively
(Figure 7.2). We chose the horizontal range as the maximum distance to accept
a trace for the geostastistical estimation of the rock properties.
99
Figure 7.1: Simplified 3D stochastic inversion workflow.
Figure 7.2: Anisotropic semivariogram model with a 0.1 nugget, a sill of 0.9 and
ranges of 500 meters and 50 meters.
100
Figure 7.3 shows the evolution of the variance model for one of the
stochastic simulation outcomes. The order selection of the seismic traces plays
an important role in the stochastic inversion process. To assure a random
selection of the trace’s order we implemented the Fisher and Yates shuffle
algorithm supported by a Markov Chain for each stochastic realization.
Therefore, any trace selection path should be independent from the previous
realizations. The upper section of Figure 7.3 shows the initial variance model, it is
clear that the closer to the Well-1 the smaller the variance. The lower variance
values are situated at the well location where the variance becomes zero. The
second, third and fourth sections (from top to bottom) represent three more
iterations that produced three new wells. Every new iteration reduced the
variance model (reducing the uncertainty) until the variance model was zero at
the fiftieth iteration.
The inversion was performed for the four seismic lithofacies defined in
Chapter 6: shales, gas sands, LSGS and brine sands. Figure 7.4 shows three
possible models generated by the proposed inversion algorithm. The total
porosity, volume of shale, P-wave velocity, bulk density and S-wave velocity
inversion outcomes are shown in figures 7.5 to 7.9, respectively. All the shown
models are probable and similar. However, they have small variations in the
depth of the seismic lithofacies interfaces, especially, in the gas sands and LSGS
interfaces where the impedance contrast is very small.
101
Figure 7.3: Variance model evolution for one of the stochastic simulation
outcomes.
Figure 7.4: Three possible stochastic models. Shales are represented by facies
1, gas sands by facies 2, LSGS by facies 3 and brine sands by facies 4.
102
Figure 7.5: Three possible stochastic total porosity models.
Figure 7.6: Three possible stochastic volume of shale models.
103
Figure 7.7: Three possible stochastic P-wave impedance models.
Figure 7.8: Three possible stochastic bulk density models.
104
Figure 7.9: Three possible stochastic S-wave velocity models.
A set of four hundred possible outcomes was generated from the seismic
section. This set of seismic inversions was used to calculate the most expecting
values of the estimated rock properties for each sample of the 2D seismic
section. As it is expected, these inverted rock property sections were smoother
versions of the original inverted outcomes.
Additionally, we generated the variance section for each rock property
inverted vertical section. The variance sections were used to better understand
the uncertainty associated with the inversion outcomes. Figure 7.10 shows the
most expected section of total porosity (upper section) and its associated
variance (lower section). Well-1 measured total porosity log is located in the trace
number 37. It can be observed that there is good agreement between the
measured data and the inverted section. However, Well-1 has a larger vertical
105
resolution than the inverted traces. As we observed in Chapter 6, the median of
the stochastic inversion outcomes tends to slightly underestimate larger porosity
values and overestimate lower porosity values. The variance section shows that
the upper gas sand has larger estimation uncertainty. This is due to the
interbedded gas-shale geometry of the upper sand. The zero variance trace
corresponds to the Well-1 location. The same phenomenon is observed in all the
rock property vertical sections; the upper sand has larger estimation errors than
the lower sand (Figures 7.11 to 7.14). In the bulk density inverted section (Figure
7.13) the variance of the upper sand and the lower sand gas intervals has larger
estimation errors than those of the LSGS and brine sand. This is caused by the
low predictability of the gas sands and LSGS fluid contact.
Figure 7.10: Most expected total porosity model generated from 400 stochastic
models.
106
Figure 7.11: Most expected volume of shale model generated from 400
stochastic models.
Figure 7.12: Most expected volume of P-wave velocity generated from 400
stochastic models.
107
Figure 7.13: Most expected volume of density generated from 400 stochastic
models.
Figure 7.14: Most expected volume of S-wave velocity generated from 400
stochastic models.
108
7.4. Convergence Criteria
A critical parameter for the proposed stochastic inversion execution is the
misfit threshold α. For this real data application we chose the L2-norm value
obtained from the synthetic tie as value of α. This selection was based on the
idea that the forward model operator is imperfect and cannot perfectly reproduce
real seismic data. Other sources of misfits, besides the forward modeling
operator, come from imperfections in the seismic data acquisition/processing and
deficiencies in the time-to-depth conversion model. It is important to point out that
the smaller α becomes, the more realizations are needed to obtain a statistically
significant number of probable models. Figure 7.15 shows 150,000 models
randomly selected from the five million realizations performed for the real
example inversion. It can be observed that only a few hundreds models look like
the studied real data set inverted.
Figure 7.15: Models randomly selected from the real dataset inversion.
109
The synthetic traces generated from the models shown in Figure 7.15 are
shown in Figure 7.16. For other shown models, the calculated seismic traces are
very noisy, without recognizable seismic interfaces. For the 150,000 traces
shown in Figure 7.16, only nine were selected as probable solutions. The
selection of α value smaller than the similarity obtained from the well log tie
would have deteriorated the inversion outcome, because the possible solutions
would have been generated from a partially sampled model space.
Figure 7.16: Comparison between real angle trace (right) and 150,000 solutions
(left) randomly selected from the five million realizations performed for the real
example inversion.
110
7.5. Algorithm Efficiency
The proposed algorithm was extensively tested in a Windows based
personal computer with two processors and eight cores. The results showed that
the amount of resources (computing time) needed by the our proposed algorithm
increase linearly (O(n)) with the number of tested models. The average running
time is 269 models/sec.
Figure 7.17: Algorithm efficiency, the proposed algorithm increase linearly with
the number of tested models.
111
6.7. Conclusions
The applicability of our proposed 3D stochastic inversion algorithm to real
case studies was demonstrated. The outcome of the inversion was a set of key
rock property sections that can be directly used for the evaluation of the
hydrocarbon potential of a given the inverted subsurface interval. Moreover, the
inversion solutions were generated directly in the depth domain. Therefore, with
our solutions, no time-to-depth is needed during the inversion process.
Results of the 3D inversion showed good vertical resolution and adequate
lateral continuity. As we observed in the 1D real data application, the real data
test of the 3D stochastic inversion shows a remarkable similarity to the available
well log data. However, the median of the inverted traces has lower vertical
variability than the Well-1 logs. Larger estimation errors were found in the upper
gas sand, due to its interbedded geometry.
The real-data tests showed that the similarity factor α is the most sensible
parameter in the proposed inversion algorithm. The selection of α value smaller
than the similarity obtained from the well log tie will deteriorate the inversion
outcome, since the possible solutions will be generated from a partially sampled
model space.
112
REFERENCES
Aki, K., and Richards, P. G., 1980, Quantitative seismology: theory and methods:
W. N. Treeman & Co. pp. 119-149.
Arpat, B. G., 2005, Sequential simulation with patterns: Ph.D. dissertation,

Stanford Univ. pp. 7-11.
Avseth, P., 2000, Combining rock physics and sedimentology for seismic
reservoir characterization in North Sea turbidite systems: Ph.D. dissertation,
Stanford Univ. pp. 14-49.
Avseth, P., Mukerji, T., and Mavko, G., 2005, Quantitative seismic interpretation:
applying rock physics tools to reduce interpretation risk: Cambridge University
Press. pp. 42-47.
Backus, G. E., 1962, Long-wave elastic anisotropy produced by horizontal

layering: J. Geophys. Res., 67, 4427-4440.
Balch, A. H., 1971, Color sonagrams - A new dimension in seismic data

interpretation: Geophysics, 36, 1074-1098.
Banchs, R. and Michelena, R., 2000, Well log estimates and confidence intervals
by using artificial neural networks: 70th Ann. Internat. Mtg,. Soc. of Expl.
Geophys., Expanded Abstracts, 1430-1432.
Batzle, M., and Wang., Z., 1992, Seismic properties of pore fluids: Geophysics,
57, 1396-1408.
Berryman, J. G., Berge, P. A., Bonner, B. P., 2002, Estimating rock porosity and
fluid saturation using only seismic velocities: Geophysics, 67, 391-404.
Bortoli, L. J., Alabert, F. A., Haas, A., and Journel, A. G., 1993, Constraining
Stochastic Images to Seismic Data, in A. Soares ed., Geostatistics Tróia 1992,
proc. 4th Inter. Geostat. Congr., Kluwer Academic Publ., 325-337.
Bosch, M., 1999, Lithologic tomography: From plural geophysical data to

lithology estimation: Journal of Geophysical Research, 104, 749–766.
Castagna, J. P., Batzle, M. L., and Eastwood, R. L., 1985, Relationships between
compressional-wave and shear-wave velocities in clastic silicate rocks:
Geophysics, 50, 571-581.
113
Castagna, J. P., Batzle, M. L., and Kan, T. K., 1993, Rock physics – The link
between rock properties and AVO response, in J. P. Castagna and M. Backus,
eds., Offset-dependent reflectivity – Theory and practice of AVO analysis: Soc. of
Expl. Geophys., 135-171.
Connolly, P., 1998, Calibration and inversion of non-zero offset seismic: 68th
Ann. Internat. Mtg,. Soc. of Expl. Geophys., Expanded Abstracts, 182-184.
Connolly, P., 1999, Elastic impedance: The Leading Edge, 18, 438-452.
Deutsch C., and Journel, A., 1998, GSLIB: Geostatistical Software Library:
Oxford University Press, 2nd edition.
Domenico, S. N., 1976, Effect of brine-gas mixture on velocity in an

unconsolidated sand reservoir: Geophysics, 41, 882-894.
Dvorkin, J., and Nur, A., 1996, Elasticity of High-Porosity Sandstones: Theory for
Two North Sea Datasets, Geophysics, 61, 1363-1370.
Dvorkin, J., Nur, A., and Yin, H., 1994, Effective Properties of Cemented
Granular Materials, Mechanics of Materials, 18, 351-366.
Fisher, A., Yates, F., 1948, Statistical tables for biological, agricultural and
medical research: Oliver & Boyd. Pp. 26-27.
Florez, J., 2005, Integrating geology, rock physics, and seismology for reservoir
quality prediction: Ph.D. dissertation, Stanford Univ. pp. 13-44.
Gardner, G., Gardner, L., and Gregory, A., 1974, Formation velocity and density
– The diagnostic basic for stratigraphic traps: Geophysics, 39, 770-780.
Gassmann, F., 1951, Über die elastizität poröser medien: Veirteljahrsschrift der
Naturforschenden Gesellschaft in Zürich, 96, 1-23.
Gonzalez, E. F., Mukerji, T. and Mavko, G., 2000, Facies identification using P-
to-P and P-to-S AVO attributes: 70th Ann. Internat. Mtg,. Soc. of Expl. Geophys.,
Expanded Abstracts, 98-101.
Gonzalez, E., 2006, Physical and quantitative Interpretation of seismic attributes

for rock and fluid interpretation: Ph.D. dissertation, Stanford Univ. pp. 49-78.
Goodway, B., Chen, T. and Downton, J., 1997, Improved AVO fluid detection and
lithology discrimination using Lame petrophysical parameters; “λρ”, “μρ”, & “λ/μ
114
fluid stack”, from P and S inversions: 67th Ann. Internat. Mtg,. Soc. of Expl.
Geophys., Expanded Abstracts, 183-186.
Grechka, V., 2008, Fluid substitution in porous and fractured solids: The non-
interaction approximation and Gassmann theory: International Journal of
Fracture, 148, 103-107.
Gray, D., 2002, Elastic inversion for Lamé parameters: 72th Ann. Internat. Mtg,.
Soc. of Expl. Geophys., Expanded Abstracts, 213-216.
Greenberg, M. L., and Castagna, J.P., 1992, Shear-wave velocity estimation in

porous rocks: Theoretical formulation, preliminary verification and applications,
Geophys. Prospect., 40, 195-209.
Guardiano, F., and Srivastava, R., 1993, Multivariate geostatistics: Beyond

bivariate moments, in A. Soares ed., Geostatistics Tróia 1992, proc. 4th Inter.
Geostat.Congr., Kluwer Academic Publ., 133-144.
Gutierrez, M., 2001, Rock physics and 3-D seismic characterization of reservoir
heterogeneities to improve recovery efficiency: Ph.D. dissertation, Stanford Univ.
Haas, A. and Dubrule, O., 1994, Geostatistical inversion - A sequential method of

stochastic reservoir modeling constrained by seismic data: First Break, 12, 561-
569.
Han, D., 1986, Effects of porosity and clay content on acoustic properties of
sandstones and unconsolidated sediments: Ph.D. thesis, Stanford University.
Harris, D. A., Lewis, J. J. M., and Wallace, D. J., 1993, The identification of
lithofacies types in geological imagery using neural networks, Conf. papers vol.,
EUROCAIPEP 93, Aberdeen.
Hashin, Z., and Shtrikman, S., 1963, A variational approach to the elastic
behavior of multiphase materials: J. Mech. Phys. Solids, 11, 127-140.
Hastings, W.K., 1970, Monte Carlo sampling methods using Markov Chains and
their applications: Biometrika, 57, 97–109.
Hearst, J. R., Nelson, P. H., Pailet, F. L., 2000, Well logging for physical
properties: McGraw-Hill, 2nd edition.
115
Hirsche, K., Boerner, S., Kalkomey, C. and Gastaldi, C., 1998, Avoiding pitfalls in
geostatistical reservoir characterization: A survival guide: The Leading Edge, 17,
493-504.
Isaaks, E., and Srivastava, R., 1989, An introduction to applied geostatistics:

Oxford University Press. pp. 278-322.
Kalkomey, C. T., 1997, Potential risks when using seismic attributes as

predictors of reservoir properties: The Leading Edge, 16, 247-251.
Landrø, M., 1999, Discrimination between pressure and fluid saturation changes
from time lapse seismic data, 69th Ann. Int. Meet., Soc. Expl. Geophys.,
Mallick, S., 2001, Hybrid inversion, elastic impedance inversion, and prestack
waveform inversion; 71st Ann. Internat. Mtg: Soc. of Expl. Geophys., 706-709.
Marion, D., 1990, Acoustical, mechanical and transport properties of sediments
and granular materials: Ph.D. thesis, Stanford University.
Matteucci, G., 1996, Seismic attribute analysis and calibration: A general

procedure and a case study: 66th Ann. Internat. Mtg,. Soc. of Expl. Geophys.,
Mavko, G., and Jizba, D., 1991, Estimation grain-scale fluid effects on velocity
dispersion in rocks: Geophysics, 56, 1940-1949.
Mavko, G., Mukerji, T., and Dvorkin, J., 1998, The rock physics handbook:
Cambridge University Press.
Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., Teller, E., 1953,
Equations of state calculations by fast computing machines: Journal of Chemical
Physics, 21, 6, 1087–1092.
Mukerji, T., Jørstad, A., Mavko, G., and Granli, J., 1998, Near and far offset
impedances: Seismic attributes for identifying lithofacies and pore fluids:
Geophysical Research Letter, 25, 4557-4560.
Mukerji, T., Avseth, P., Mavko, G, Takahashi, I., and Gonzalez, E. F., 2001,
Statistical rock physics: Combining rock physics, information theory, and
geostatistics to reduce uncertainty in seismic reservoir characterization: The
Leading Edge, 20, 313-319.
116
Oldenburg, D. W., Scheuer, T., and Levy, S., 1983, Recovery of the acoustic
impedance from reflection seismograms: Geophysics, 48, 1318-1337.
Ostrander, W. J., 1984, Plane-wave reflection coefficients for gas sands at non-
normal angles of incidence: Geophysics, 49, 1637-1648.
Papoulis, A., 1984, Probability, random variables, and stochastic processes:

McGraw-Hill.
Possato, S., Saito, M., Curtis, M. P., and Martinez, R. D., 1983, Interpretation of
three-dimensional seismic attributes contributes to stratigraphic analysis of
Pampo oil field: 53rd Ann. Internat. Mtg,. Soc. of Expl. Geophys., Expanded
Abstracts, session: S16.2.
Raymer, L. L., E. R. Hunt, and J. S. Gardner, 1980, An improved sonic transit

time-to-porosity transform: 21st Annual Logging Symposium Transaction, Society
of Professional Well Log Analysts, Paper P.
Ruiz, F., 2009, Porous grain model and equivalent elastic medium approach for
predicting effective elastic properties of sedimentary rocks: Ph.D. dissertation,
Stanford Univ.
Russell, B. H., 1988, Introduction to seismic inversion methods: Soc. of Expl.

Geophys. Course notes from SEG Continuing Education course.
Sengupta, M., 2000, Integrating rock physics and flow simulation to reduce
uncertainties in seismic reservoir monitoring: Ph.D. dissertation, Stanford Univ.
Shuey, R. T., 1985, A simplification of the Zoeppritz equations: Geophysics, 50,

609-614.
Sheriff, R., 1989, Geophysical Methods: Prentice Hall.
Takahashi, I., 2000, Quantifying information and uncertainty of rock property

estimation from seismic data: Ph.D. dissertation, Stanford Univ.
Taner, M. T., Koehler, F., and Sheriff, R. E., 1979, Complex seismic trace
analysis: Geophysics, 44, 1041-1063.
Tarantola, A., 2005, Inverse problem theory and methods for model parameter
estimation: Society for Industrial and Applied Mathematics.
Thomsen, L., 1985, Biot-consistent elastic moduli of porous rocks: Low-

frequency limit: Geophysics, 50, 2797-2807.
117
Tonn, R., 1992, Statistical approach to correlate reservoir parameters and 3-D
seismic attributes, 62nd Ann. Internat. Mtg,. Soc. of Expl. Geophys., Expanded
Abstracts, 272-274.
Wu, Y., 2000, Estimation of gas saturation using P-to-S converted waves; 70th
Ann. Internat. Mtg,. Soc. of Expl. Geophys., Expanded Abstracts, 158-161.
Wyllie, M. R., Gregory, A. R., and Gardner, L. W., 1956: Elastic wave velocities in
heterogeneous and porous media, Geophysics, 21, 41-70.
Yin, H., Mavko, G., and Nur, A., 1993, Critical porosity; a physical boundary in
poroelasticity: 34th U. S. symposium on Rock mechanics, in Haimson, B. C., ed.,
Rock mechanics in the 1990s, Oxford-New York, International Journal of Rock
Mechanics and Mining Sciences & Geomechanics, Abstracts, 30, 805-808.
Zeng, H., Backus, M. M., Barrow, K. T. and Tyler, N., 1996, Facies mapping from
three-dimensional seismic data: Potential and guidelines from a Tertiary
sandstone-shale sequence model, Powderhorn Field, Calhoun County, Texas:
AAPG Bull., 80, 16-46.
118

Tesis Doctorado Carlos Cobos

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Tesis Doctorado Carlos Cobos

Transféré par

Droits d'auteur :

Formats disponibles

MARKOV CHAIN STOCHASTIC SEISMIC INVERSION FOR

ROCK PROPERTIES ESTIMATION

The Faculty of the Department Earth and Atmospheric Sciences

of the Requirements for the Degree

ROCK PROPERTIES ESTIMATION

Dr. John Castagna, Chairman

Dr. Evgeny M. Chesnokov, Member

Dr. Aibing Li, Member

Dr. Franklin Ruiz, Member

Dean, College of Natural Sciences and

I would like to express my gratitude to John Castagna for being an

support I would not have been able to complete this dissertation.

Ezequiel Gonzalez, Andres Mantilla, Per Avseth, Mario Gutierrez, Isao

Takahashi, Manuel Gonzalez and Marcelo Benabentos. I am also very grateful to

they have given me all these years.

I have been fortunate to share my personal time with a number of great

I have infinite gratitude to my parents Gracia and Carlos, my grandfather

Finally, I dedicate this dissertation to my wife Carmen and my children

that made possible the culmination of my Ph.D.

ROCK PROPERTIES ESTIMATION

The Faculty of the Department Earth and Atmospheric Sciences

of the Requirements for the Degree

Our purpose is to introduce a stochastic seismic inversion algorithm based

on the Markov Chain Monte Carlo Simulation. Conventional inversion algorithms

characterize the hydrocarbon potential of a given subsurface interval. In contrast,

the inversion scheme presented here generates a set of possible combinations of

rock properties that can explain seismic amplitude responses in terms of

results consist of many equally probable rock property models as an alternative

to multiple elastic property scenarios. Therefore, no post facto elastic-to-rock-

properties conversion is needed. The method might be used either in exploratory

areas or hydrocarbon field development. In exploratory areas, the stochastic rock

physics inversion can support the evaluation for hydrocarbon potential

considering the effects of reservoir properties on seismic signatures for different

uncertainties and risk. In field development areas, stochastic seismic inversion

3D seismic response; it can be used to constrain possible reservoir models used

for hydrocarbon reserve estimation and reservoir production simulation. The

probabilistic inversion algorithm was successfully tested on a set of real seismic

properties for hydrocarbon exploration, such as total porosity and reservoir

inversion technique to accurately predict the rock properties of reservoir seismic

lithofacies, even for seismically thin layers.

Acknowledgments ...................................................................................... iii

Table 2.1: Probability associated to the possible outcomes of random process. .. 9

Conventional seismic reservoir characterization (SRC) techniques were

seismic responses were relatively easy to identify. Early reservoir

characterization workflows were usually based on direct hydrocarbon indicator

(DHI) identification techniques centered on poststack seismic amplitude analysis

shallow high porosity reservoirs with significantly lower acoustic impedance

compared to the surrounding rocks. The associated seismic signatures of these

hydrocarbon-filled high porosity reservoirs can be anomalous high amplitude

reflections called “bright spots”.

Nowadays, conventional seismic-reservoir-characterization techniques are

hydrocarbons are located in deeper and more-complex reservoirs. These new

hydrocarbon plays are characterized by low-porosity and low-permeability

future seismic-reservoir characterization goal is to predict rock properties, such

resolution. Currently, seismic inversion resolution is still low in new

exploration/development challenges and the improvement of seismic derived

elastic parameters is paramount for the application of reservoir properties

prediction. Techniques such as stochastic inversion have been initially developed

in an attempt to obtain from seismic data quantitative information about

subsurface rock properties on a very detailed scale.

The main limitation of current seismic-inversion algorithms is that the

characterizing hydrocarbon potential. To accurately characterize the oil potential

of a possible of subsurface reservoir, we need to know key rock properties, such